betaFeedback

WorksComputer Science

Corridor-Based Monitoring and Emergency Braking Decision for Automated Vehicle Safety

Abstract

This paper addresses emergency-braking decisions for a redundant safety channel in automated vehicles. The channel is responsible for identifying hazards from sensor data and determining whether braking is necessary, given the constrained computational resources of embedded hardware. We present an emergency-braking approach that uses a predicted motion corridor to filter \ac{lidar} measurements and then decide whether the ego vehicle must brake. \ac{pmcm} constructs a curvature-aware boundary from the planned path and vehicle geometry so that the monitored region contains the full swept body on both straight and curved paths. \ac{gdeta} then decides whether the ego vehicle must command an emergency brake for objects inside this boundary, scoring geometric occupancy and temporal urgency independently. Hard gates permit braking when either source is conclusive; soft-OR arbitration combines partial evidence when neither gate is active. Consecutive-scan confirmation reduces sensitivity to transient measurement noise. Corridor coverage is evaluated in controlled collision simulations and recorded campus driving. Braking decisions are evaluated in Monte Carlo simulations of lead-vehicle braking and transverse-crossing encounters.

Introduction

Modern Advanced Driver Assistance System (ADAS) stacks increasingly rely on Artificial Intelligence (AI), especially deep neural networks, to raise nominal perception performance [1]. These models are not easy to inspect and may produce detections that are hard to explain [2], or they may degrade on rare, long-tail traffic situations that are poorly represented in the training data [3, 4]. A second channel, with its own sensor, decision rules, and authority to stop the vehicle, is an established redundancy when the main autonomy stack misses a hazard or leaves too little time to avoid a collision [5, 6].

This paper studies such a second channel that runs on low-cost hardware, such as microcontrollers, and that can support a safety argument. The monitor takes a trajectory planned by the High-Level Autonomy System (HLAS) and raw scans from a single LiDAR, and at each scan decides whether an emergency stop is required within a bounded monitoring region.

For such resource-limited monitoring systems, two questions follow. First, how can the monitoring area be restricted to the traffic participants that may collide with the ego vehicle as it follows the planned path? Second, once those participants occupy that area, how can the available object measurements and the known ego motion be used to decide an emergency brake command?

Last-feasible-brake rules determine when to stop based on objects that have already been tracked [7, 8], but they generally assume that geometric relevance has already been established. Where an explicit monitoring region is used, it is often a coarse keep-out mask [6, 9]. On a tight turn, such a mask may be too narrow along the outer edge, so an obstacle enters the monitored region only after the available braking distance has been consumed. The same mask may be too wide beside the path, so curbs or parked vehicles appear relevant.

We use the same predicted motion corridor to retain LiDAR returns within the swept body and to decide whether the ego vehicle must command an emergency brake for the resulting objects. The geometric basis is PMCM, a curvature-aware monitoring boundary. The planned rear-axle path is expressed in a curvilinear frame, the four body corners are propagated along it, and sliding extrema of their tracks form an envelope that contains the swept body at any curvature. GDETA then scores geometric occupancy and temporal urgency independently on this envelope, so either source can command a stop when its evidence is conclusive.

Keeping the two evidence sources separate lets GDETA use the information the tracker already has. Object position is available before velocity is reliable, so the geometric evidence can command braking once the object occupies the emergency envelope for the required number of scans, or immediately if the object is already close and blocks enough of the corridor. After a valid motion estimate exists, the temporal evidence checks whether the object will still overlap the ego width when the ego reaches that longitudinal position, and whether braking has become urgent. The temporal condition must persist over consecutive scans before a command is issued, so a brief predicted overlap or one noisy measurement does not stop the vehicle.

Evaluation starts with corridor coverage and then isolates the brake rule. A controlled collision series places a static obstacle at the outer corridor edge across a range of speeds and turning radii. Among the three masks considered, the PMCM envelope is the only one that both detects the obstacle in every trial and prevents collisions in all tested cases after braking. The parallelogram mask detects the obstacle too late, while the two-point strip often fails to detect it. The GDETA threat assessment decision rule is then evaluated independently in two Monte Carlo campaigns. An established lead-vehicle case tests temporal assessment with the geometric gate suppressed by a confirmed track. A transverse-crossing case tests geometric intervention during the first scans after a sudden corridor entry, before a reliable motion estimate and tracker lock are available.

This article extends our previous conference publication [10], which introduced the curvature-aware PMCM method and evaluated it in a case study. The present article reformulates PMCM and extends it with:

  1. GDETA, a deterministic threat-assessment rule that scores geometric occupancy and temporal urgency independently, admits a hard gate on either source, and uses soft-OR arbitration when no gate applies.

  2. An evaluation that separates corridor coverage from braking decisions and examines GDETA in established lead-vehicle braking and transverse-crossing encounters.

Related Work

This section reviews emergency braking and independent safety channels, corridor and envelope monitoring, and threat-assessment metrics for an embedded brake rule.

Emergency Braking and Independent Safety Channels

Autonomous Emergency Braking (AEB) systems estimate collision risk from nearby objects and apply a brake command when necessary [11]. Production algorithms are mainly longitudinal and work on an object list: last-feasible braking when a collision is unavoidable given vehicle dimensions and orientation [7], model-based remaining-maneuver sets that also include steering and acceleration [8], and later timing rules that delay full brake for comfort [12]. They are built as a primary ADAS function. The input is a tracked object list; the planned automated trajectory is not part of the monitor.

Emergency intervention also appears as an independent safety channel beside a higher-level stack. Hanselaar et al. [5] propose a multi-channel Safety Shell that arbitrates among redundant functions under a Safety of the Intended Functionality (SOTIF) argument. Pek et al. [13, 14] execute an intended trajectory only while a collision-free fail-safe trajectory remains available under set-based occupancies of other road users, using both steering and braking. A LiDAR-based AEB system for a heavy vehicle uses sparse flash returns [15]. It connects the sensor to the brake command and selects relevant targets, still as a primary AEB function.

Althoff and Dolan verify candidate vehicle motion online against reachable occupancies [16]. Gruber and Althoff return conservative verification results under varying computational load with an anytime refinement scheme [17]. Magdici and Althoff keep a verified fail-safe motion behind the intended trajectory [18], and Schürmann et al. formally check whether planned motions remain dynamically drivable [19]. vom Dorff et al. separate a safety observer from the nominal planner and buffer short trajectories that end in a safe stop [20]. These methods check planned motion at runtime and keep a safe stop available.

Corridor and Envelope Monitoring

Geometric monitoring regions can serve as the collision-check region for braking. Althoff et al. [21] predict the stochastic road occupancy of surrounding vehicles and integrate crash probability along a given ego trajectory, so the monitored region is a time-varying occupancy density. Lee and Kum [22] and Shen et al. [23] construct a predictive occupancy map from the relative position, velocity and acceleration of nearby traffic, score a discrete set of acceleration candidates on that map, and trigger collision mitigation from the same occupancy.

On curved roads, a fixed or straight-ahead region assigns the wrong objects to the path. Lee et al. [24] convert host and target states into curvilinear, road-aligned coordinates before computing the relative distance and time-to-collision that trigger AEB. Han et al. [25] reshape the region of interest online from steering angle and longitudinal velocity, and adapt its range under a worst-case road geometry without an explicit lane or curvature estimate. That adaptation reduces false detections of roadside objects such as guardrails and raises detection of genuine in-path threats. Collision checking can also use a free-space contour: Gao et al. [26] represent instantaneous free space as a Doppler-deformable polygon and check collisions against that polygon instead of a dense occupancy grid.

Threat Assessment

Dahl et al. [27] and Westhofen et al. [28] group collision metrics into single-behavior scalars (Time-to-Collision (TTC), Time Headway (TH), Time-to-Brake (TTB), Brake Threat Number (BTN), Time-to-React (TTR)), optimization and reachability, probabilistic frameworks, and learning-based scores. The monitor uses the braking-only scalars TTC, TTB and BTN, which an embedded rule can evaluate deterministically.

Jansson and Gustafsson [29] give a Bayesian collision-avoidance decision from required acceleration, the quantity later called BTN. Brännström et al. [8] compute an analytical remaining set of steer, brake and accelerate maneuvers for a rectangular host and a polygonal object, and brake when that set is empty. Kaempchen et al. [7] trigger when a collision is unavoidable over all physically possible trajectories. Gelso and Sjöberg [30] use TTB and BTN together, with road geometry to predict lead-vehicle deceleration, to suppress false alarms at junctions on euroFOT data.

Reachable sets and constraint satisfaction bound last-chance timing more tightly than a scalar TTB. Falcone et al. [31] treat threat as constraint satisfaction and receding-horizon safe sets for road departure. Sönges, Koschi and Althoff [32] over-approximate TTR with reachable sets; an empty set means no evasive maneuver remains. Sánchez Morales et al. [33] evaluate millions of ego/object pose hypotheses on a Drive PX 2 in about 21 ms.

Xu, Zhao and Wang [34] synthesize TTC, time-headway and a Time-to-Front term after Interacting Multiple Model prediction, then search an Markov Decision Process for a trajectory; the fused score is a planner cost. Responsibility-Sensitive Safety (RSS) [35] specifies formal longitudinal and lateral duty-of-care distances. The geometric emergency envelope here is the planned corridor limited to a kinematic delay-plus-braking depth.

Hillenbrand et al. organize collision mitigation as multilevel situation assessment and decision-making, and make the intervention trade-offs explicit [36]. Eidehall and Petersson estimate collision threat by Monte Carlo sampling of uncertain road scenes [37]. Schreier, Willert and Adamy couple maneuver recognition, trajectory prediction and criticality assessment in arbitrary road layouts [38]. Kim and Kum combine lane-conditioned motion probabilities with trajectory-pair TTC into a collision-risk map over candidate ego paths [39]. These methods use prediction distributions, maneuver hypotheses or candidate trajectories.

System Overview

NeFiMMORE is the dedicated near-field monitoring system of the Munich Mobility Research Campus (MORE) experimental vehicle. It uses its own 3D LiDAR and an embedded host. The hardware platform architecture is given in [6, 40].

System Description

Figure 1 shows the autonomous vehicle technical stack. The platform is split into a main stack and a fail-safe stack, each with a system layer, a sensor layer, and an actuator layer. The main stack is the HLAS: a high-performance computer with roof-mounted high-resolution LiDAR, camera, radar, and GNSS (Figure 2). It plans a collision-free path and commands steering and the powertrain. The fail-safe stack is NeFiMMORE: an embedded host and a low-mounted dedicated LiDAR, with authority to command the service brake. The stacks share two messages. The HLAS sends the planned trajectory. The monitor returns a warning when a relevant object occupies the predicted motion corridor, and a brake command when Sect. 4.2 requires a stop. NeFiMMORE is the redundant last-resort channel: the HLAS remains the first actor on a conflict it can observe in time, and the monitor commands a stop for a sudden corridor hazard that the roof suite cannot process in time.

Figure 1. Automated vehicle technical stack. The fail-safe stack is Near-Field Monitoring System MORE (NeFiMMORE) with a dedicated LiDAR and brake authority. The main stack is the HLAS with the full sensor suite, steering, and powertrain. The only runtime coupling is the planned trajectory toward the monitor and a warning plus brake request back to the vehicle [10].

Experimental vehicle sensor setup

Figure 2. Experimental vehicle sensor setup. Our NeFiMMORE system uses a low-mounted front 3D LiDAR to perceive the near-field environment.

The roof LiDAR can detect far traffic, but it may still miss a near-field object in the dead corner of the roof LiDAR. From the planned path and the current speed, the monitor builds a predicted motion corridor whose length covers reaction plus braking and whose width follows the body sweep [10, 41]. A low-mounted front LiDAR watches the volume near the ground inside that corridor.

Figure 3 illustrates the monitoring corridor in representative scenes. The red region marks the keep-out area. A speed-dependent rectangle is sufficient on a straight road [41]; on a tight turn, however, the corridor must follow the geometric boundary of the ego vehicle. § 4.1 describes this construction.

System Pipeline

Figure 4 shows the runtime pipeline from corridor monitoring to the brake decision. Two inputs enter in parallel: the planned ego track from the HLAS and the raw sensor input from the dedicated LiDAR. A prefilter and a main filter discard LiDAR samples outside the region of interest, which reduces the memory and compute load of processing the raw point cloud. The retained points are grouped into objects, associated across scans, and passed to the risk analyzer. The boundary finder and the main filter implement PMCM (Sect. 4.1). The risk analyzer implements GDETA (Sect. 4.2).

Figure 4. Overview of the runtime processing pipeline implemented in NeFiMMORE.

Figure 3. Predicted motion corridor (marked in red) in representative near-field scenes. Longitudinal depth follows the current speed; lateral bounds follow the planned body sweep [10].

The boundary finder builds the monitoring corridor whose length follows the current speed and whose width follows the predicted body sweep. The prefilter crops the LiDAR sweep to a near-field volume in the sensor frame, so returns outside that volume never reach the corridor test. The main filter then maps the corridor into the LiDAR frame and retains the LiDAR returns that lie inside it.

Clustering groups those points by Euclidean region growing [42]. Object tracking associates clusters across scans by Hungarian assignment of centroids [43], keeps a stable identifier, and estimates motion. The footprint and the motion are then projected onto the planned path, so each track is expressed in the corridor coordinates that GDETA uses. The risk analyzer receives this object list together with the ego motion. Geometric occupancy of the emergency envelope and temporal braking urgency are scored independently, and either source may command a stop. The pipeline output is a binary hard-brake request.

Proposed Methodology

Curve-aware Monitoring using PMCM

PMCM maps the planned driving path of the HLAS and the ego vehicle geometry onto a curvature-aware region of interest, then retains only those LiDAR returns that fall inside this region. The region follows the predicted body sweep, so it remains valid when path curvature changes. On a tight urban turn, a fixed-width or parallelogram mask is either too small, so it misses the outer part of the swept path, or too large, so it treats sidewalk, curb, or parked objects beside the turn as if the ego vehicle would drive through them. Clustering and tracking operate on corridor-resident points, and GDETA in Sect. 4.2 assesses only objects that already lie inside this monitored set.

The procedure has three stages. First, a motion corridor is obtained from sliding extrema of the vehicle-corner tracks. Second, that corridor is reduced to a braking-horizon, sensor-visible monitoring boundary. Third, each return is tested against azimuth-local triangles of this boundary.

Motion-corridor Envelope

A curvilinear frame is attached to the predicted reference track of the rear-axle midpoint, shown in Figure 5. Let γm:[0,L]→R2\gamma_m : [0,L] \to\mathbb{R}^{2} be a C2C^{2} planar curve parameterized by arc length ss,

Predicted reference path in a curvilinear frame with motion-corridor envelope

Figure 5. Predicted reference path γm(s)\gamma_m(s) in a curvilinear frame, together with the motion-corridor envelope obtained from the vehicle-corner tracks [10].

γm(s)=(xm(s)ym(s)),(1)\gamma_m(s) = \begin{pmatrix} x_m(s) \\ y_m(s) \end{pmatrix}, \tag*{(1)}

with unit tangent tm(s)=γm′(s)t_m(s) = \gamma_m'(s) and leftward unit normal

nm(s)=(−ym′(s)xm′(s)).(2)n_m(s) = \begin{pmatrix} -y_m'(s) \\ x_m'(s) \end{pmatrix}. \tag*{(2)}

A ground-plane point then has coordinates (s,r)(s,r) through p(s,r)=γm(s)+nm(s) r\mathbf{p}(s,r) = \gamma_m(s) + n_m(s)\,r. The same corridor frame is later used by GDETA, where the lateral coordinate is written dd.

The vehicle body R\mathcal{R} is modeled as a rectangle

R={(a,b)∈R2:−lr≤a≤lf, ∣b∣≤w/2},(3)\mathcal{R} = \{(a,b) \in\mathbb{R}^{2} : -l_r \leq a \leq l_f,\ |b| \leq w/2\}, \tag*{(3)}

where lrl_r and lfl_f are the rearward and forward overhangs from the vehicle baselink and ww is the outer width. At pose s0s_0, a body point (a,b)∈R(a,b) \in\mathcal{R} maps to pcar(s0,a,b)=γm(s0)+atm(s0)+bnm(s0)\mathbf{p}_{\mathrm{car}}(s_0,a,b) = \gamma_m(s_0) + a t_m(s_0) + b n_m(s_0). The swept set S=⋃s0∈[0,L](γm(s0)+Rs0)S = \bigcup_{s_0 \in[0,L]}(\gamma_m(s_0) + \mathcal{R}_{s_0}) is the region that must be monitored. Lateral bounds rL(s)r_L(s) and rR(s)r_R(s) are required to satisfy

S⊆{(s,r):rR(s)≤r≤rL(s)}.(4)S \subseteq\{(s,r) : r_R(s) \leq r \leq r_L(s)\}. \tag*{(4)}

Rather than forming this Minkowski sweep explicitly, the method tracks the four corner trajectories {(sil,rilf)}\{(s_i^{\mathrm{l}},r_i^{\mathrm{lf}})\}, {(sil,rilr)}\{(s_i^{\mathrm{l}},r_i^{\mathrm{lr}})\}, {(sir,rirf)}\{(s_i^{\mathrm{r}},r_i^{\mathrm{rf}})\}, {(sir,rirr)}\{(s_i^{\mathrm{r}},r_i^{\mathrm{rr}})\}, where the superscripts denote the left front, left rear, right front, and right rear corners. The tracks are sampled on a common grid sk=kΔss_k=k\Delta s, where Δs\Delta s is the sampling interval. At sks_k the body occupies the interval [sk−lr,sk+lf][s_k-l_r,s_k+l_f]. With forward window length F=⌈lf/Δs⌉F=\lceil l_f/\Delta s\rceil and backward window length R=⌈lr/Δs⌉R=\lceil l_r/\Delta s\rceil, the left and right envelopes are the sliding extrema

rL(sk)=max⁡{max⁡j∈[k,k+F]rlf(sj), max⁡j∈[k−R,k]rlr(sj)},rR(sk)=min⁡{min⁡j∈[k,k+F]rrf(sj), min⁡j∈[k−R,k]rrr(sj)}.(5)\begin{aligned} r_L(s_k)&=\max\left\{\max_{j\in[k,k+F]}r_{\mathrm{lf}}(s_j),\ \max_{j\in[k-R,k]}r_{\mathrm{lr}}(s_j)\right\},\\ r_R(s_k)&=\min\left\{\min_{j\in[k,k+F]}r_{\mathrm{rf}}(s_j),\ \min_{j\in[k-R,k]}r_{\mathrm{rr}}(s_j)\right\}. \tag*{(5)} \end{aligned}

Each sliding extremum is linear in the number of samples. The resulting polylines, drawn in red in Figure 5, are the motion corridor used by all subsequent stages.

Visible Monitoring Boundary

The full envelope is larger than the region that a front-center LiDAR can observe before a stop. An operational monitoring boundary is therefore required to respect the braking distance, to contain only vertices that are visible from the sensor, and to use the available Field of View (FOV). Computation is performed in the LiDAR frame, whose xx-axis points forward. The three reduction stages are shown in Figure 6 and are described next.

Construction of the visible monitoring boundary from a predicted motion corridor

Figure 6. Construction of the visible monitoring boundary from a predicted motion corridor: (a) boundary preconstruction; (b) visible reduction; (c) occlusion filtering [10].

Boundary preconstruction.

Let PirP_i^r and PjlP_j^l denote the discrete vertices along the right and left corridor boundaries from Eq. (5), now expressed in the LiDAR frame. Only samples with x≥0x\geq0 are kept; vertices behind the sensor are ignored. If a side has no discrete point at x=0x=0, that intersection with the lateral axis is interpolated so the polygon is closed at the sensor. Let Δk=∥(xk+1,yk+1)−(xk,yk)∥2\Delta_k=\left\|(x_{k+1},y_{k+1})-(x_k,y_k)\right\|_2 be successive segment lengths along the predicted front-LiDAR path and Ls=∑k=0s−1ΔkL_s=\sum_{k=0}^{s-1}\Delta_k the cumulative length. The stopping index $s satisfies Ls≤db<Ls+1L_s \le d_b < L_{s+1}, where dbd_b is the braking distance. The stopping pose vms\mathbf{v}^{s}_{m} is interpolated between vs\mathbf{v}_{s} and vs+1\mathbf{v}_{s+1} so that the path length equals dbd_b, then projected onto both sides, giving vrr\mathbf{v}^{r}_{r} and vrl\mathbf{v}^{l}_{r}. The preconstructed set

Sp={Pir:x≥0, i=0,…,s−1}∪{vrs}∪{Pjl:x≥0, j=0,…,s−1}∪{vls}(6)\mathcal{S}^{p} = \{\mathcal{P}^{r}_{i}: x \ge0,\ i=0,\ldots,s-1\}\cup\{\mathbf{v}^{s}_{r}\}\cup\{\mathcal{P}^{l}_{j}: x \ge0,\ j=0,\ldots,s-1\}\cup\{\mathbf{v}^{s}_{l}\} \tag*{(6)}

is the primary forward, braking-limited corridor (green circles connected by line segments in the left panel of Figure 6).

Visible reduction.

This step walks each ordered chain and retains only vertices that remain extremal in azimuth from the sensor origin o\mathbf{o}, collecting S^r\hat{\mathcal{S}}^{r} and S^l\hat{\mathcal{S}}^{l}. With α(v)=atan2⁡(y,x)+π/2\alpha(\mathbf{v})=\operatorname{atan2}(y,x)+\pi/2, the right accumulator stores the current maximum azimuth and the left accumulator the current minimum. Self-occluded samples do not update these extrema: in the center panel of Figure 6, v11r\mathbf{v}^{r}_{11}, v12r\mathbf{v}^{r}_{12}, and the vertices lying behind v13l\mathbf{v}^{l}_{13} are excluded.

Occlusion filtering.

This step removes candidates that are hidden by the opposite corridor. In the right panel of Figure 6, v17r\mathbf{v}^{r}_{17} is occluded by v13l\mathbf{v}^{l}_{13} and is discarded. Let S^trimr\hat{\mathcal{S}}^{r}_{\mathrm{trim}} and S^triml\hat{\mathcal{S}}^{l}_{\mathrm{trim}} be the vertices of S^r\hat{\mathcal{S}}^{r} and S^l\hat{\mathcal{S}}^{l} that remain after this opposite-corridor test. The surviving vertices of both sides are merged into a single sequence ordered by increasing azimuth,

(v1,…,vM):=sort⁡α(S^trimr∪S^triml),(7)(\mathbf{v}_{1},\ldots,\mathbf{v}_{M}) := \operatorname{sort}_{\alpha}\left(\hat{\mathcal{S}}^{r}_{\mathrm{trim}}\cup\hat{\mathcal{S}}^{l}_{\mathrm{trim}}\right), \tag*{(7)}

which is the open polygonal chain used for point inclusion in the next step.

Azimuth-window Point Filter

Consecutive vertices vivi+1\mathbf{v}_i\mathbf{v}_{i+1} of Eq. (7) form a local window. A 3D LiDAR return is first projected onto the 2D plane. The resulting pj\mathbf{p}_j is retained if its azimuth lies in [α(vi),α(vi+1))[\alpha(\mathbf{v}_i),\alpha(\mathbf{v}_{i+1})) and it belongs to the closed triangle △(o,vi,vi+1)\triangle(\mathbf{o},\mathbf{v}_i,\mathbf{v}_{i+1}). Membership is decided by consistent orientation of the three edge cross products

s1=(vi+1−o)×(pj−o),s2=(vi+1−vi)×(pj−vi),s3=(o−vi+1)×(pj−vi+1).(8)\begin{aligned} s_1 &= (\mathbf{v}_{i+1}-\mathbf{o}) \times(\mathbf{p}_j-\mathbf{o}),\\ s_2 &= (\mathbf{v}_{i+1}-\mathbf{v}_i) \times(\mathbf{p}_j-\mathbf{v}_i),\\ s_3 &= (\mathbf{o}-\mathbf{v}_{i+1}) \times(\mathbf{p}_j-\mathbf{v}_{i+1}). \tag*{(8)} \end{aligned}

The orientation test holds when s1,s2,s3s_1,s_2,s_3 are all non-negative or all non-positive:

(s1≥0∧s2≥0∧s3≥0)∨(s1≤0∧s2≤0∧s3≤0)(9)(s_1 \ge0 \land s_2 \ge0 \land s_3 \ge0) \lor(s_1 \le0 \land s_2 \le0 \land s_3 \le0) \tag*{(9)}

Figure 7 shows a return pj\mathbf{p}_j inside the window and a return pj′\mathbf{p}'_j that fails it. The orientation test is expensive if applied to every 3D return, so a candidate is rejected unless it lies no farther than the more distant of the two window vertices

Figure 7. Local azimuth window of two successive visible vertices. Point pj\mathbf{p}_j lies inside the filtering triangle; pj′\mathbf{p}'_j lies outside the admissible range [10].

∥pj−o∥2≤max⁡{∥vi−o∥2,∥vi+1−o∥2}.(10)\lVert\mathbf{p}_j-\mathbf{o}\rVert_2 \le\max\{\lVert\mathbf{v}_i-\mathbf{o}\rVert_2,\lVert\mathbf{v}_{i+1}-\mathbf{o}\rVert_2\}. \tag*{(10)}

In Figure 7, pj′\mathbf{p}'_j violates Eq. (10) for the window vi′vi+1′\mathbf{v}'_i\mathbf{v}'_{i+1}. The filter may be applied to a preprocessed cloud or to the ordered raw packet stream. Points that remain inside the corridor are passed to clustering, tracking, and then to GDETA.

Gated Dual-Evidence Threat Assessment

GDETA decides, from objects already inside the monitored motion corridor, whether the ego vehicle must command an emergency brake. The rule is deterministic and intended for embedded execution. Geometric occupancy and temporal urgency are scored independently; neither score is an input to the other. Hard gates let either source command immediate braking. Soft scores in [0,1][0,1] are fused only if neither hard gate is active. The procedure consists of: (i) a relevance gate; (ii) geometric evidence scoring; (iii) a geometric emergency gate; (iv) temporal evidence scoring based on braking-oriented criticality measures and predicted corridor occupancy at arrival; (v) an Last-Feasible-Brake (LFB) gate; (vi) multi-object soft arbitration; and (vii) the final brake decision.

Coordinate frame and inputs.

Threat assessment is performed in the corridor-aligned frame of Sect. 4.1, with longitudinal arc-length coordinate ss along the ego reference path and lateral offset dd (Figure 5). Let O\mathcal{O} be the set of tracked objects that already lie inside the monitored corridor and are supplied to GDETA. For each object i∈Oi \in\mathcal{O}, the upstream tracking module supplies the occupied footprint BiB_i in the corridor frame, the nearest clearance sinears_i^{\mathrm{near}}, the stable-track flag FiF_i, the lateral center did_i, the longitudinal speed vs,iv_{s,i}, the lateral speed vd,iv_{d,i}, and the measured object width wiw_i. Envelope-occupancy persistence nihazn_i^{\mathrm{haz}} and soft-condition persistence nisoftn_i^{\mathrm{soft}} are maintained by GDETA from its own previous output on the same track identifier, matching the consecutive-count pattern of the LFB gate. The Time-to-Disappear (TTD) used later is computed inside GDETA from this body state.

Normalization.

All soft scores are clipped to the unit interval by the saturation operator

sat⁡(x)=min⁡(max⁡(x,0),1),(11)\operatorname{sat}(x)=\min(\max(x,0),1), \tag*{(11)}

where x∈Rx \in\mathbb{R} is an arbitrary scalar. A linear map onto [0,1][0,1] uses two thresholds: x0x_0 is the onset, at which the score leaves zero, and x1x_1 is the saturation, at which the score reaches one. If danger increases with xx, then x0<x1x_0 < x_1 and

R=sat⁡(x−x0x1−x0).(12)R=\operatorname{sat}\left(\frac{x-x_0}{x_1-x_0}\right). \tag*{(12)}

If danger increases as xx decreases, then x0>x1x_0 > x_1 and

R=sat⁡(x0−xx0−x1).(13)R=\operatorname{sat}\left(\frac{x_0-x}{x_0-x_1}\right). \tag*{(13)}

Hence R=0R=0 on the safe side of x0x_0, R=1R=1 on the critical side of x1x_1, and RR interpolates linearly in between.

Relevance gate.

From the input set O\mathcal{O}, the relevance gate retains the tracks that enter threat assessment. Let γi∈{0,1}\gamma_i \in\{0,1\} be the ground-connection flag of object ii and δi=max⁡(ℓimeas,wimeas)\delta_i=\max(\ell_i^{\mathrm{meas}},w_i^{\mathrm{meas}}) its observed spatial extent, the larger of the measured longitudinal and lateral footprint sizes. Two class-conditional minima are used, with δmin⁡f>δmin⁡g>\delta_{\min}^{f}>\delta_{\min}^{g}> 00: δmin⁡g\delta_{\min}^{g} for ground-connected objects and δmin⁡f\delta_{\min}^{f} for floating objects. The relevant object set Orel\mathcal{O}_{\mathrm{rel}} collects those members of O\mathcal{O} that meet the matching size floor,

Orel={i∈O:(γi∧δi≥δmin⁡g)∨(¬γi∧δi≥δmin⁡f)}.(14)\mathcal{O}_{\mathrm{rel}} = \{i \in\mathcal{O} : (\gamma_i \land\delta_i \ge\delta_{\min}^{g}) \lor(\neg\gamma_i \land\delta_i \ge\delta_{\min}^{f})\}. \tag*{(14)}

Subsequent scoring and gates operate on Orel\mathcal{O}_{\mathrm{rel}}.

Geometric evidence.

For each relevant object i∈Oreli \in\mathcal{O}_{\mathrm{rel}}, a proximity score and a blocking score are defined by

Rprox,i=sat⁡(sact−sinearΔs),(15)R_{\mathrm{prox},i} = \operatorname{sat}\left(\frac{s_{\mathrm{act}} - s_i^{\mathrm{near}}}{\Delta s}\right), \tag*{(15)}
Rblock,i=sat⁡(wiovwsat),(16)R_{\mathrm{block},i} = \operatorname{sat}\left(\frac{w_i^{\mathrm{ov}}}{w_{\mathrm{sat}}}\right), \tag*{(16)}

where sinears_i^{\mathrm{near}} is the nearest longitudinal clearance from the ego front reference to object ii along the corridor, sacts_{\mathrm{act}} is the longitudinal activation distance at which geometric concern begins, Δs>0\Delta s > 0 is the longitudinal normalization span, wiovw_i^{\mathrm{ov}} is the lateral overlap width of object ii with the monitoring corridor, and wsat>0w_{\mathrm{sat}} > 0 is the overlap width at which the blocking score saturates to one. The geometric evidence score is then

Rgeo,i=Rprox,iRblock,i∈[0,1].(17)R_{\mathrm{geo},i} = R_{\mathrm{prox},i}R_{\mathrm{block},i} \in[0,1]. \tag*{(17)}
Geometric emergency gate.

The emergency envelope is the monitored corridor of Sect. 4.1, restricted to the braking-limited longitudinal interval [0,semg][0, s_{\mathrm{emg}}]:

Eemg={(s,d):0≤s≤semg, dR(s)≤d≤dL(s)},(18)E_{\mathrm{emg}} = \{(s,d) : 0 \le s \le s_{\mathrm{emg}},\ d_R(s) \le d \le d_L(s)\}, \tag*{(18)}

with

semg=vegoτsys+vego22aavail+sbuf(19)s_{\mathrm{emg}} = v_{\mathrm{ego}}\tau_{\mathrm{sys}} + \frac{v_{\mathrm{ego}}^2}{2a_{\mathrm{avail}}} + s_{\mathrm{buf}} \tag*{(19)}

and with dL(s)d_L(s), dR(s)d_R(s) the left and right corridor bounds at arc length ss. Here vegov_{\mathrm{ego}} is the ego longitudinal speed, τsys\tau_{\mathrm{sys}} the command-path time reserve used in Eq. (19), aavail>0a_{\mathrm{avail}} > 0 the available emergency deceleration, and sbufs_{\mathrm{buf}} a longitudinal safety buffer. Envelope occupancy Hi∈{0,1}H_i \in\{0,1\} is one when the tracked occupied footprint BiB_i intersects EemgE_{\mathrm{emg}} and the visible width meets a minimum,

Hi=(Bi∩Eemg≠∅)∧(wiov≥wmin⁡),(20)H_i = (B_i \cap E_{\mathrm{emg}} \ne\varnothing) \land(w_i^{\mathrm{ov}} \ge w_{\min}), \tag*{(20)}

Envelope-occupancy confirmation is the binary flag Cihaz∈{0,1}C_i^{\mathrm{haz}} \in\{0,1\}. GDETA keeps a consecutive envelope-occupancy count nihazn_i^{\mathrm{haz}} on the same track identifier: the count increments while Hi=1H_i = 1 and resets otherwise, including when geometry is invalid. The flag is one when that count reaches the confirmation length NCgeoN_C^{\mathrm{geo}},

Cihaz=1{nihaz≥NCgeo}.(21)C_i^{\mathrm{haz}} = 1\{n_i^{\mathrm{haz}} \ge N_C^{\mathrm{geo}}\}. \tag*{(21)}

Immediate geometric braking is reserved for envelope occupancy that is already both wide and close,

Qistrong=(wiov≥wblock)∧(sinear≤simm),(22)Q_i^{\mathrm{strong}} = (w_i^{\mathrm{ov}} \ge w_{\mathrm{block}}) \land(s_i^{\mathrm{near}} \le s_{\mathrm{imm}}), \tag*{(22)}

with 0<wmin⁡<wblock0 < w_{\min} < w_{\mathrm{block}}, and simm>0s_{\mathrm{imm}} > 0 a near-contact gap. A confirmed tracker lock suppresses the geometric hard gate. Immediate geometric braking is commanded when

Gigeo=¬Fi∧Hi∧(Cihaz∨Qistrong),(23)G_i^{\mathrm{geo}} = \neg F_i \land H_i \land\left(C_i^{\mathrm{haz}} \lor Q_i^{\mathrm{strong}}\right), \tag*{(23)}

where Fi∈{0,1}F_i \in\{0,1\} is the stable-track flag.

Temporal evidence.

Temporal evidence is computed independently of the geometric evidence. Longitudinal closing uses the constant-acceleration relative-gap model

gi(t)=sgap,i−vrel,it−12arel,it2,(24)g_i(t) = s_{\mathrm{gap},i} - v_{\mathrm{rel},i}t - \frac{1}{2}a_{\mathrm{rel},i}t^2, \tag*{(24)}

where gi(t)g_i(t) is the predicted longitudinal clearance at time t≥0t \ge0, sgap,is_{\mathrm{gap},i} is the current longitudinal gap to object ii, vrel,i=vego−vobj,iv_{\mathrm{rel},i} = v_{\mathrm{ego}} - v_{\mathrm{obj},i} is the relative closing speed (positive when closing), and arel,i=aego−aobj,ia_{\mathrm{rel},i} = a_{\mathrm{ego}} - a_{\mathrm{obj},i} is the relative longitudinal acceleration. The TTC is the smallest positive root of gi(t)=0g_i(t) = 0. From these kinematics, the required braking magnitude areq,ia_{\mathrm{req},i} yields the BTN

BTNi=areq,iaavail,(25)\mathrm{BTN}_i = \frac{a_{\mathrm{req},i}}{a_{\mathrm{avail}}}, \tag*{(25)}

where areq,i≥0a_{\mathrm{req},i} \ge0 is the minimum ego braking deceleration needed immediately to avoid a predicted collision. The TTB is the latest admissible delay before applying aavaila_{\mathrm{avail}} that still avoids collision. If no collision is predicted, TTCi=TTBi=∞\mathrm{TTC}_i = \mathrm{TTB}_i = \infty and BTNi=0\mathrm{BTN}_i = 0. Normalized temporal sub-scores RTTB,i,RBTN,i∈[0,1]R_{\mathrm{TTB},i}, R_{\mathrm{BTN},i} \in[0,1] are obtained by mapping TTBi\mathrm{TTB}_i with Eq. (13) and BTNi\mathrm{BTN}_i with Eq. (12) between warning (x0x_0) and critical (x1x_1) thresholds. Those longitudinal metrics measure closing urgency.

The front-arrival time tiarrt_i^{\mathrm{arr}} is the first instant at which Eq. (24) reaches the object’s current longitudinal position. Three outcomes are possible:

tiarr={0if sgap,i≤0,TTCif such a t exists,undefinedif sgap,i>0 and no such t exists.(26)t_i^{\mathrm{arr}} = \begin{cases} 0 & \text{if } s_{\mathrm{gap},i} \le0,\\ \mathrm{TTC} & \text{if such a } t \text{ exists},\\ \text{undefined} & \text{if } s_{\mathrm{gap},i} > 0 \text{ and no such } t \text{ exists.} \end{cases} \tag*{(26)}

The corridor width at position ss is the span of the bounds already used in Eq. (18),

wc(s)=dL(s)−dR(s).(27)w_{\mathrm{c}}(s) = d_{\mathrm{L}}(s) - d_{\mathrm{R}}(s). \tag*{(27)}

The lateral exit offset ri=12(wc+wi)r_i = \frac{1}{2}(w_{\mathrm{c}} + w_i) is the distance the object center must travel from the corridor centerline to leave the corridor. Arrival occupation Rocc,iR_{\mathrm{occ},i} is one if the object center is still inside the corridor, by at least the leftover-stay margin, when the front arrives.

Rocc,i=1{∣di+vd,itiarr∣≤ri−∣vd,i∣τocc}.(28)R_{\mathrm{occ},i} = 1\left\{\left|d_i + v_{d,i}t_i^{\mathrm{arr}}\right| \le r_i - |v_{d,i}|\tau_{\mathrm{occ}}\right\}. \tag*{(28)}

The margin τocc\tau_{\mathrm{occ}} is the minimum leftover stay counted as occupied. If tiarrt_i^{\mathrm{arr}} is undefined, Rocc,i=0R_{\mathrm{occ},i} = 0. Leftover stay is the time the object remains in the corridor after front arrival. The object leaves the corridor at

TTDi=sup⁡{t≥0:∣di+vd,it∣≤ri},(29)\mathrm{TTD}_i = \sup\left\{t \ge0 : \left|d_i + v_{d,i}t\right| \le r_i\right\}, \tag*{(29)}

with TTDi=∞\mathrm{TTD}_i = \infty if the object remains in the band for all future time. Then τstay,i=TTDi−tiarr\tau_{\mathrm{stay},i} = \mathrm{TTD}_i - t_i^{\mathrm{arr}} when the body occupies the raw band at tiarrt_i^{\mathrm{arr}}. The fused temporal evidence score is

Rtemp,i=Rocc,i(1−(1−RTTB,i)(1−RBTN,i))∈[0,1].(30)R_{\mathrm{temp},i} = R_{\mathrm{occ},i}\left(1 - (1 - R_{\mathrm{TTB},i})(1 - R_{\mathrm{BTN},i})\right) \in[0,1]. \tag*{(30)}
lfb gate.

The flag ViV_i admits the LFB gate only when the object is still in the corridor at front arrival and remains there for at least the leftover-stay threshold Tm>τoccT_m > \tau_{\mathrm{occ}}:

Vi=1{Rocc,i=1∧τstay,i≥Tm}.(31)V_i = 1\left\{R_{\mathrm{occ},i} = 1 \land\tau_{\mathrm{stay},i} \ge T_m\right\}. \tag*{(31)}

A scan is urgent only when Vi=1V_i = 1 and either remaining brake delay or required deceleration is already critical,

RiLFB,now=Vi∧((TTBi≤τact+τbuf)∨(BTNi≥θbtn)),(32)\mathcal{R}_i^{\mathrm{LFB,now}} = V_i \land\left((\mathrm{TTB}_i \le\tau_{\mathrm{act}} + \tau_{\mathrm{buf}}) \lor(\mathrm{BTN}_i \ge\theta_{\mathrm{btn}})\right), \tag*{(32)}

where τact\tau_{\mathrm{act}} is the brake actuation delay, τbuf\tau_{\mathrm{buf}} is a temporal safety buffer, and θbtn=BTNcrit=1\theta_{\mathrm{btn}} = \mathrm{BTN}_{\mathrm{crit}} = 1 is the saturation threshold of required versus available deceleration. Either longitudinal clause is sufficient; both are gated by ViV_i. The consecutive count niriskn_i^{\mathrm{risk}} increments while RiLFB,now\mathcal{R}_i^{\mathrm{LFB,now}} holds for the same identifier and resets otherwise. Immediate LFB braking is commanded when

GiLFB=1{nirisk≥NCLFB},(33)G_i^{\mathrm{LFB}} = 1\left\{n_i^{\mathrm{risk}} \ge N_{\mathrm{C}}^{\mathrm{LFB}}\right\}, \tag*{(33)}

so the gate confirms over NCLFBN_{\mathrm{C}}^{\mathrm{LFB}} scans.

Soft arbitration and final decision.

For objects on which neither hard gate has fired, a per-object Risk Index (RI) is formed by soft-OR fusion,

RIi=1−(1−Rgeo,i)(1−Rtemp,i),(34)\mathrm{RI}_i = 1 - (1 - R_{\mathrm{geo},i})(1 - R_{\mathrm{temp},i}), \tag*{(34)}

and the scene threat index is taken as the dominant-object value RIscene=max⁡iRIi\mathrm{RI}_{\mathrm{scene}} = \max_i \mathrm{RI}_i. The instantaneous soft condition on object ii requires arrival occupation, a positive temporal score, and an index at or above the threshold,

Sinow=(Rocc,i=1)∧(Rtemp,i>0)∧(RIi≥θRI),(35)S_i^{\mathrm{now}} = (R_{\mathrm{occ},i} = 1) \land(R_{\mathrm{temp},i} > 0) \land(\mathrm{RI}_i \ge\theta_{\mathrm{RI}}), \tag*{(35)}

where θRI∈(0,1]\theta_{\mathrm{RI}} \in(0,1] is a soft braking threshold. The consecutive count nisoftn_i^{\mathrm{soft}} increments while SinowS_i^{\mathrm{now}} holds for the same identifier and resets otherwise. Let i⋆i^\star be the dominant object in the residual soft set. The ego brake command is issued when

Brake⇐(∃i:Gigeo=1)∨(∃i:GiLFB=1)∨(ni⋆soft≥NCri).(36)\mathrm{Brake} \Leftarrow\left(\exists i : G_i^{\mathrm{geo}} = 1\right) \lor\left(\exists i : G_i^{\mathrm{LFB}} = 1\right) \lor\left(n_{i^\star}^{\mathrm{soft}} \ge N_C^{\mathrm{ri}}\right). \tag*{(36)}

Experimental Evaluation and Validation

Evaluation of PMCM

Section 4.1 constructs a path-conforming swept-body envelope and retains those LiDAR returns that fall inside it. That construction is the geometric precondition for GDETA, which consumes tracked objects already inside the monitored set. The experiments below ask whether, relative to simpler masks, the PMCM envelope covers the predicted body sweep tightly enough that a static obstacle on the corridor edge is detected in time.

Controlled Collision Simulation

Experimental setup.

The study is performed in Gazebo with a sensor model of a Velodyne VLP-16 that matches the experimental vehicle. A static cylinder of radius 0.35 m and height 0.6 m is placed on the outer edge of the driving corridor. In each trial the ego vehicle holds a constant speed on either a circular path or a straight road. Speeds range from 1 km/h to 30 km/h in steps of 2 km/h. The path radii are {6,8,10,12,14,16,18,20,22,24,26,∞}\{6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, \infty\} m, where an infinite radius denotes the straight-road case. The simulated vehicle uses the same body dimensions as the experimental platform. For every trial the first detected obstacle position is stored, and it is recorded whether a subsequent emergency stop still avoids contact. All designed encounters are collision cases, so the comparison uses the False Negative Rate (FNR), the True Positive Rate (TPR), and the residual Collision Rate (CR) after a commanded stop. Figure 8 shows one such encounter, together with the parallelogram mask, the corridor mask, and the first-detection location marked pos. A.

Bird’s-eye view of a simulated corridor-edge collision case

Figure 8. Bird’s-eye view of a simulated corridor-edge collision case. The parallelogram mask, the PMCM corridor, and the first-detection location pos. A are indicated [10].

Reference filters.

Two alternative masks are evaluated on the same trials. The first is a parallelogram filter that adapts a quadrilateral keep-out region to the instantaneous driving state. The ratio of parallelogram area to PMCM corridor area was computed for each radius. The ratio increases toward unity as the path straightens, and it quantifies how much corridor area the parallelogram loses on tight turns. The second baseline, two-point corridor filtering, traces only the two rear vehicle corners along the predicted path and treats the resulting strip as the monitor. It has a corridor-like outline, yet it does not apply the sliding-extrema envelope of Eq. (5).

Detection timing and collision outcome.

Table 1 summarizes the three masks. Both the parallelogram filter and PMCM attain a FNR of 0% and a TPR of 100%: the cylinder is eventually seen in every trial. Their residual CR values differ. The parallelogram mask leaves a collision in 36.1% of the trials, whereas PMCM leaves none. The recorded first-detection poses show why. Under the parallelogram mask the cylinder often becomes visible only at pos. A in

MeasureParallelogramTwo-point corridorPMCM
FNR0%57.74%0%
TPR100%42.26%100%
CR36.1%1.03%0%

Table 1. Comparison of FNR, TPR, and residual CR for three monitoring masks on the same controlled collision series. Adapted from [10].

Figure 8, which already lies inside the braking distance dbd_b. A stop commanded at that instant cannot prevent impact. As curvature increases, the parallelogram occupies a shrinking fraction of the true swept corridor, so an obstacle on the outer edge enters the mask too late even though the object is eventually detected. On large radii the same ratio approaches one, and the parallelogram becomes an acceptable stand-in for the path-conforming envelope.

The two-point strip attains a residual CR of 1.03%, close to that of PMCM, but it misses the cylinder in 57.74% of the trials. Most of those misses occur at radii of 16 m or less. Two rear-corner generators cannot follow the outer body sweep on a tight turn, so a corridor-edge obstacle may lie outside the strip for the entire approach. That is a different failure mode from the parallelogram case: the two-point mask often never sees the object, whereas the parallelogram sees it too late. PMCM avoids both modes because the sliding-extrema envelope of Sect. 4.1 is built from the vehicle geometry itself.

Recorded Campus Driving

The same envelope is next applied to recorded campus runs, which include urban-like turns, sensor noise, ego-motion fluctuation, and localization drift that the Gazebo series does not reproduce. Figure 9 shows one snippet in a camera view and a corresponding bird’s-eye overlay. The left and right corridor polylines update continuously along the predicted path, remain aligned with the vehicle body sweep, and follow changes in turning radius without switching mask type. The recording is used to check that the online boundary stays aligned under those disturbances.

Online PMCM corridor on recorded campus driving, shown in a camera view and a bird’s-eye overlay

Figure 9. Online PMCM corridor on recorded campus driving, shown in a camera view and a bird’s-eye overlay [10].

Evaluation of GDETA

GDETA consumes tracked objects that are already inside the monitored set O\mathcal{O}. This subsection tests the decision rule of Sect. 4.2 on object-level states already expressed in the corridor frame. Two cases are studied. An established lead has a stable track, so the geometric hard gate is suppressed and the trial tests temporal and soft braking. A transverse crossing first appears at the corridor boundary; this is the sudden corridor entry the monitor is meant to cover, and the trial tests the geometric gate and its later handover to temporal assessment.

Evaluation Design

The decision logic is evaluated in a Monte Carlo simulation. Each trial is split into three parallel worlds.

Ground-truth world.

This world propagates the ego and one object at 100 Hz. In the lead case both bodies use constant acceleration along the path: the lead cruises, then brakes after a sampled time. In the crossing case the target holds along-path speed at zero and moves through the corridor at constant lateral speed. Two noise-free rollouts supply labels. The no-brake rollout keeps constant ego speed and marks the trial dangerous if the bodies meet. The oracle is a second rollout in which the ego applies the same delayed first-order brake (a 0.2 s dead time and a 0.15 s rise toward 7.5 m/s27.5\ \mathrm{m/s^2}) at the first instant a causal reaction is defined. For lead braking that instant is a physics event: the sampled time when the lead starts to brake. For crossing, the target is invisible until its body overlaps the corridor, and the monitor only looks at 10 Hz scans, so the oracle latches on the first scan at which that overlap exists. A remaining oracle collision is labeled unavoidable and is excluded from the avoidable-collision denominator.

gdeta input world.

This world converts the current ground-truth state into corridor-aligned object metadata. Independent Gaussian noise is added only here, on those decision inputs, to emulate imperfect tracking. Both campaigns use standard deviations of 0.5 m, 0.5 m/s, and 0.5 m/s20.5\ \mathrm{m/s^2} on gap, longitudinal speed, and longitudinal acceleration. Crossing also uses 0.1 m/s on lateral speed. Its configuration records a 0.05 m lateral-position deviation. For the lead case, Fi=1F_i = 1 from the first scan. For the crossing case, motion state is invalid on the first two visible scans, and FiF_i stays 0 for three visible scans and then switches to 1 to indicate a valid track. At every 10 Hz scan the method returns a brake command together with the geometric and temporal scores and the hard-gate flags of Sect. 4.2.

Evaluation world.

This world closes the loop around GDETA on the same ground-truth plant. After the latch, ego deceleration follows the actuator model above. The run stops at impact or mutual standstill. The closed-loop outcome is judged against the two ground-truth counterfactuals: the no-brake rollout defines whether the sample is dangerous, and the oracle flags collisions that remain after that earliest command.

Monte Carlo Sampling

The campaign covers the two cases listed in Table 2. Continuous quantities are sampled independently from uniform distributions. The speed range of 3–30 km/h matches the intended near-field operation, while the lead-vehicle deceleration range of [3.0, 7.5] m/s² covers moderate through emergency braking. The GDETA settings are fixed: NCego=2N_C^{\mathrm{ego}} = 2, NCLFB=3N_C^{\mathrm{LFB}} = 3, NCri=2N_C^{\mathrm{ri}} = 2, θRI=0.89\theta_{\mathrm{RI}} = 0.89, aavail=7.5a_{\mathrm{avail}} = 7.5 m/s², Twarn=0.6T_{\mathrm{warn}} = 0.6 s, Tcrit=0.08T_{\mathrm{crit}} = 0.08 s, BTNwarn=0.6\mathrm{BTN}_{\mathrm{warn}} = 0.6, and BTNcrit=1.0\mathrm{BTN}_{\mathrm{crit}} = 1.0.

PropertyEstablished lead brakingTransverse crossing
Trials / seed200,000 / 10200,000 / 10
Horizon60 s12 s
Ego speed3–30 km/h3–30 km/h
Target body3.5 × 2.0 m vehicle0.5 × 0.4 m target
Target motion0–30 km/h; 3–7.5 m/s² brakingvs=0v_s = 0; ∣vd∣=0.5\lvert v_d \rvert= 0.5–3 m/s
Encounter timinggap 0.5–15 m; brake onset 0–5 scenterline-arrival offset −3–3 s
Flag FiF_i1 from the first scan0 for three visible scans, then 1

Table 2. Object-level Monte Carlo scene design. Continuous quantities are sampled independently from uniform distributions.

In the lead case, the leading vehicle remains centered with zero lateral velocity, so Rocc,i=1R_{\mathrm{occ},i} = 1, τstay,i=∞\tau_{\mathrm{stay},i} = \infty, and Vi=1V_i = 1 throughout a typical trial. The fixed Fi=1F_i = 1 suppresses GigeoG_i^{\mathrm{geo}}, and any command must come from the LFB gate or soft arbitration. In the crossing case the target starts just outside the corridor, left or right with equal probability, and has no along-path speed. Its initial gap is computed from a sampled arrival offset in [−3, 3] s: offset zero means the ego front reaches the target’s fixed path position at the same instant the target reaches the corridor centerline. The geometric gate may fire while Fi=0F_i = 0 and motion is still invalid; once Fi=1F_i = 1, Eq. (23) suppresses that gate and later scans use the temporal and soft paths.

Results and Discussion

Table 3 reports the simulation results. Brake-command recall is computed among real collisions (True Positive (TP) and False Negative (FN)) that the causal oracle can avoid. A false positive is a command in a no-brake non-collision trial.

MetricsLead brakingTransverse crossing
Trials200,000200,000
Unavoidable cases59,69093,840
FNR0.05%24.20%
brake-command recall99.95%75.80%
precision99.99%97.05%
False Positive Rate (FPR)4/18 (22.22%)799/71,497 (1.12%)

Table 3. GDETA Monte Carlo results.

In the lead-braking campaign, GDETA identifies nearly all avoidable longitudinal conflicts, missing 75 of 140,292 cases. Across all lead trials, including cases classified by the oracle as unavoidable, the first command comes from soft arbitration in 199,484 cases and from the LFB gate in 406 cases. With the stable-track flag fixed at Fi=1F_i = 1, none comes from the geometric gate. The campaign contains only 18 no-brake non-collisions, so the observed 4/18 FPR is a thin sample of nuisance braking. The missed cases are low-speed approaches to a stopped lead vehicle, for which both TTB and BTN remain below their activation thresholds until contact. The temporal urgency maps therefore miss that low-speed contact condition.

In the transverse-crossing campaign the two hard gates and soft arbitration all fire. The method commands braking in 26,276 of the 34,663 avoidable collision cases. Among the 39,475 crossing commands, the first command is geometric-only in 10,148 cases, temporal-only in 7,017, and soft-only in 22,310. Newly visible targets can therefore reach the geometric gate before FiF_i suppresses it, and later scans can use the temporal or soft rule. Of the 71,497 no-brake non-collision trials, 799 receive a command, so the FPR is 1.12%. The avoidable-collision FNR is 24.20%. An emergency stop on a non-collision interrupts the planned motion, so the consecutive-scan gates and θRI\theta_{\mathrm{RI}} keep the false-alarm rate low. The resulting FN/False Positive (FP) operating point is that of a redundant last-resort brake: a crossing that remains visible for several scans is a case the HLAS is expected to treat first, and NeFiMMORE is reserved for a sudden corridor intrusion that the roof suite cannot process in time. No intervention induces a collision in either campaign.

After three visible scans FiF_i becomes 1 and the geometric gate closes, so later braking needs consecutive urgent frames from the LFB gate or from soft arbitration. A 0.4 m pedestrian reaches that urgency only when the gap is already small. When a command does appear, it is on average 0.35 s before the no-brake impact, with a mean TTB of 0.048 s.

Limitations

NeFiMMORE is a redundant last-resort monitor of the body sweep predicted from the HLAS path. Objects outside the motion corridor are excluded from its decision scope. Restricting the monitored region is a trade-off between computational demand and coverage. The method responds to hazards that enter the corridor, and its decisions rely on a single 3D LiDAR. Consecutive-scan confirmation in GDETA is set to keep nuisance emergency stops low; the HLAS remains the first actor on crossings it can observe in time.

The geometric evidence score in GDETA varies linearly with the front gap. A nonlinear mapping may be more appropriate when the decision should become more sensitive as the gap approaches zero.

The simulation considers one object at a time: either a longitudinally braking lead vehicle or a lateral target moving at constant speed. The longitudinal model assumes constant relative acceleration, while the crossing model assumes constant lateral speed. Multi-object interactions and lane-entry scenarios are left to future work.

Conclusion

This article combines Predictive Motion Corridor Monitoring (PMCM) and Gated Dual-Evidence Threat Assessment (GDETA) in one Near-Field Monitoring System MORE (NeFiMMORE) system. PMCM constructs a curvature-aware monitoring boundary from the path supplied by the High-Level Autonomy System (HLAS) and the vehicle geometry, and uses that boundary to filter LiDAR returns. GDETA then decides whether the ego vehicle must command an emergency brake for tracked objects in the same corridor. Geometric occupancy and temporal urgency stay separate: either hard gate can issue a brake command, and soft arbitration combines the two scores when both gates remain inactive.

PMCM is evaluated by placing an obstacle at the outer corridor edge over a range of speeds and path radii. The proposed boundary detects every tested obstacle; the comparison masks either see it too late or miss it. Object-level Monte Carlo trials then test GDETA in established lead-braking and transverse-crossing cases. Lead-vehicle brake-command recall is 99.95%. In the crossing case all three decision mechanisms fire, and the false-alarm rate among non-collisions is 1.12%. The HLAS remains the first actor on crossings it can observe in time; NeFiMMORE is reserved for a sudden corridor intrusion.

Acknowledgements

This research is funded by dtec.bw – Digitalization and Technology Research Center of the Bundeswehr, which we gratefully acknowledge. dtec.bw is funded by the European Union – NextGenerationEU. We also acknowledge financial support from the FORscience Research Fund of the Universität der Bundeswehr München. The authors thank Matthias Gerdts and Thomas Rottmann for valuable discussions, and the anonymous reviewers for their feedback.

Declarations

Competing interests

The authors declare that they have no competing financial or other interests that could influence the research or its outcomes.

Ethical Approval

This research did not involve human participants, animals, or sensitive data, so ethical approval was not required.

Not applicable. This study did not involve human participants.

Not applicable. This study did not involve human participants.

References

  1. [1]Shivanna, V.M., Guo, J.-I.: Object detection, recognition, and tracking algorithms for ADAS—a study on recent trends. Sensors 24(1), 249 (2023) https://doi.org/10.3390/s24010249
  2. [2]Zablocki, É., Ben-Younes, H., Pérez, P., Cord, M.: Explainability of deep vision-based autonomous driving systems: Review and challenges. International Journal of Computer Vision 130(10), 2425–2452 (2022) https://doi.org/10.1007/s11263-022-01657-x
  3. [3]Albrecht, S., Hirsch, A., Raafatnia, S., Woehrle, M.: Deep learning safety concerns in automated driving perception. IEEE Transactions on Intelligent Vehicles 10(2), 1455–1466 (2025) https://doi.org/10.1109/TIV.2024.3428415
  4. [4]Liu, H.X., Feng, S.: Curse of rarity for autonomous vehicles. Nature Communications 15(1) (2024) https://doi.org/10.1038/s41467-024-49194-0
  5. [5]Hanselaar, C.A.J., Silvas, E., Terechko, A., Heemels, W.P.M.H.: The safety shell: An architecture to handle functional insufficiencies in automated driving. IEEE Transactions on Intelligent Transportation Systems 25(7), 7522–7540 (2024) https://doi.org/10.1109/TITS.2024.3352829
  6. [6]Pan, J., Sotiriadis, P., Englberger, F.: Independent near-field monitoring: A novel approach to improve active safety in autonomous vehicles. In: Proc. IEEE Intelligent Vehicles Symposium (IV), Jeju Island, Korea, pp. 730–736 (2024). https://doi.org/10.1109/IVS55156.2024.10588775
  7. [7]Kaempchen, N., Schiele, B., Dietmayer, K.: Situation assessment of an autonomous emergency brake for arbitrary vehicle-to-vehicle collision scenarios. IEEE Transactions on Intelligent Transportation Systems 10(4), 678–687 (2009) https://doi.org/10.1109/TITS.2009.2026452
  8. [8]Brännström, M., Coelingh, E., Sjöberg, J.: Model-based threat assessment for avoiding arbitrary vehicle collisions. IEEE Transactions on Intelligent Transportation Systems 11(3), 658–669 (2010) https://doi.org/10.1109/TITS.2010.2048314
  9. [9]Gutjahr, B., Werling, M.: Automatic collision avoidance during parking and maneuvering — an optimal control approach. In: Proc. IEEE Intelligent Vehicles Symposium (IV), Dearborn, MI, USA, pp. 636–641 (2014). https://doi.org/10.1109/IVS.2014.6856575
  10. [10]Pan, J., Mansharamani, M., Sotiriadis, P., Nenchev, V., Englberger, F.: Enhancing functional reliability of autonomous vehicle safety monitoring in curves. In: Proc. 12th International Conference on Vehicle Technology and Intelligent Transport Systems (VEHITS) (2026). https://doi.org/10.5220/0014812400004030
  11. [11]Yang, L., Yang, Y., Wu, G., Zhao, X., Fang, S., Liao, X., Wang, R., Zhang, M.: A systematic review of autonomous emergency braking system: Impact factor, technology, and performance evaluation. Journal of Advanced Transportation 2022, 1–13 (2022) https://doi.org/10.1155/2022/1188089
  12. [12]Han, S., Choi, K.H., Kim, K.-S.: Adaptive safety distance for collision avoidance based on a full braking activation parameter. IEEE Transactions on Intelligent Transportation Systems 25(9), 12657–12668 (2024) https://doi.org/10.1109/TITS.2024.3386165
  13. [13]Pek, C., Manzinger, S., Koschi, M., Althoff, M.: Using online verification to prevent autonomous vehicles from causing accidents. Nature Machine Intelligence 2(9), 518–528 (2020) https://doi.org/10.1038/s42256-020-0225-y
  14. [14]Pek, C., Althoff, M.: Fail-safe motion planning for online verification of autonomous vehicles using convex optimization. IEEE Transactions on Robotics 37(3), 798–814 (2021) https://doi.org/10.1109/TRO.2020.3036624
  15. [15]Kumar, V., Subramanian, S.C., Rajamani, R.: Autonomous emergency braking of a heavy road vehicle using a low-density flash lidar. IEEE Transactions on Vehicular Technology 73(2), 1879–1889 (2024) https://doi.org/10.1109/TVT.2023.3314534
  16. [16]Althoff, M., Dolan, J.M.: Online verification of automated road vehicles using reachability analysis. IEEE Transactions on Robotics 30(4), 903–918 (2014) https://doi.org/10.1109/TRO.2014.2312453
  17. [17]Gruber, F., Althoff, M.: Anytime safety verification of autonomous vehicles. In: 2018 21st International Conference on Intelligent Transportation Systems (ITSC), pp. 1708–1714 (2018). https://doi.org/10.1109/ITSC.2018.8569950
  18. [18]Magdici, S., Althoff, M.: Fail-safe motion planning of autonomous vehicles. In: 2016 IEEE 19th International Conference on Intelligent Transportation Systems (ITSC), pp. 452–458 (2016). https://doi.org/10.1109/ITSC.2016.7795594
  19. [19]Schürmann, B., Heß, D., Eilbrecht, J., Stursberg, O., Köster, F., Althoff, M.: Ensuring drivability of planned motions from simple models using formal methods. In: ITSC 2017 (2017). https://elib.dlr.de/111912/
  20. [20]Dorff, S.v., Bö ddeker, B., Kneissl, M., Fränzle, M.: A fail-safe architecture for automated driving. In: 2020 Design, Automation & Test in Europe Conference & Exhibition (DATE), pp. 828–833 (2020). https://doi.org/10.23919/DATE48585.2020.9116283
  21. [21]Althoff, M., Stursberg, O., Buss, M.: Model-based probabilistic collision detection in autonomous driving. IEEE Transactions on Intelligent Transportation Systems 10(2), 299–310 (2009) https://doi.org/10.1109/TITS.2009.2018966
  22. [22]Lee, K., Kum, D.: Collision avoidance/mitigation system: Motion planning of autonomous vehicle via predictive occupancy map. IEEE Access 7, 52846–52857 (2019) https://doi.org/10.1109/ACCESS.2019.2912067
  23. [23]Shen, D., Chen, Y., Li, L., Chien, S.: Collision-free path planning for automated vehicles risk assessment via predictive occupancy map. In: 2020 IEEE Intelligent Vehicles Symposium (IV), pp. 985–991 (2020). https://doi.org/10.1109/IV47402.2020.9304720
  24. [24]Lee, J., Kim, G., Kim, B.: Study on the improvement of a collision avoidance system for curves. Applied Sciences 9(24), 5380 (2019) https://doi.org/10.3390/app9245380
  25. [25]Han, S., Choi, M., Jang, B., Choi, K.H., Kim, K.-S.: Adaptive roi for collision warning mitigation based on road geometry and kinematics. IEEE Sensors Journal 25(3), 5499–5509 (2025) https://doi.org/10.1109/JSEN.2024.3512947
  26. [26]Gao, X., Ding, S., Dasari, H.R.: Deformable radar polygon: A lightweight and predictive occupancy representation for short-range collision avoidance. IEEE Sensors Journal 24, 26104–26114 (2024) https://doi.org/10.1109/JSEN.2024.3422148
  27. [27]Dahl, J., Campos, G.R., Olsson, C., Fredriksson, J.: Collision avoidance: A literature review on threat-assessment techniques. IEEE Transactions on Intelligent Vehicles 4(1), 101–113 (2019) https://doi.org/10.1109/TIV.2018.2886682
  28. [28]Westhofen, L., Neurohr, C., Koopmann, T., Butz, M., Schütt, B., Utesch, F., Neurohr, B., Gutenkunst, C., Böde, E.: Criticality metrics for automated driving: A review and suitability analysis of the state of the art. Archives of Computational Methods in Engineering 30(1), 1–35 (2023) https://doi.org/10.1007/s11831-022-09788-7
  29. [29]Jansson, J., Gustafsson, F.: A framework and automotive application of collision avoidance decision making. Automatica 44(9), 2347–2351 (2008) https://doi.org/10.1016/j.automatica.2008.01.016
  30. [30]Gelso, E.R., Sjöberg, J.: Consistent threat assessment in rear-end near-crashes using BTN and TTB metrics, road information and naturalistic traffic data. IEEE Intelligent Transportation Systems Magazine 9(1), 74–89 (2017) https://doi.org/10.1109/MITS.2016.2550012
  31. [31]Falcone, P., Ali, M., Sjöberg, J.: Predictive threat assessment via reachability analysis and set invariance theory. IEEE Transactions on Intelligent Transportation Systems 12(4), 1352–1361 (2011) https://doi.org/10.1109/TITS.2011.2158210
  32. [32]Söntges, S., Koschi, M., Althoff, M.: Worst-case analysis of the time-to-react using reachable sets. In: Proc. IEEE Intelligent Vehicles Symposium (IV), Changshu, China, pp. 1891–1897 (2018). https://doi.org/10.1109/IVS.2018.8500709
  33. [33]Sánchez Morales, E., Membarth, R., Gaull, A., Slussalek, P., Dirndorfer, T., Kammenhuber, A., Lauer, C., Botsch, M.: Parallel multi-hypothesis algorithm for criticality estimation in traffic and collision avoidance. In: Proc. IEEE Intelligent Vehicles Symposium (IV), Paris, France, pp. 2164–2171 (2019). https://doi.org/10.1109/IVS.2019.8814015
  34. [34]Xu, C., Zhao, W., Wang, C.: An integrated threat assessment algorithm for decision-making of autonomous driving vehicles. IEEE Transactions on Intelligent Transportation Systems 21(6), 2510–2521 (2020) https://doi.org/10.1109/TITS.2019.2919865
  35. [35]Shaley-Shwartz, S., Shammah, S., Shashua, A.: On a Formal Model of Safe and Scalable Self-driving Cars (2017). http://arxiv.org/abs/1708.06374
  36. [36]Hillenbrand, J., Spieker, A.M., Kroschel, K.: A multilevel collision mitigation approach—its situation assessment, decision making, and performance tradeoffs. IEEE Transactions on Intelligent Transportation Systems 7(4), 528–540 (2006) https://doi.org/10.1109/TITS.2006.883115
  37. [37]Eidehall, A., Petersson, L.: Statistical threat assessment for general road scenes using Monte Carlo sampling. IEEE Transactions on Intelligent Transportation Systems 9(1), 137–147 (2008) https://doi.org/10.1109/TITS.2007.909241
  38. [38]Schreier, M., Willert, V., Adamy, J.: An integrated approach to maneuver-based trajectory prediction and criticality assessment in arbitrary road environments. IEEE Transactions on Intelligent Transportation Systems 17(10), 2751–2766 (2016) https://doi.org/10.1109/TITS.2016.2522507
  39. [39]Kim, J., Kum, D.: Collision risk assessment algorithm via lane-based probabilistic motion prediction of surrounding vehicles. IEEE Transactions on Intelligent Transportation Systems 19(9), 2965–2976 (2018) https://doi.org/10.1109/TITS.2017.2768318
  40. [40]Pan, J., Sotiriadis, P., Englberger, F.: Collision avoidance of autonomous driving at low speed in the near field of vehicle. International Journal of Emerging Engineering and Technology 2(1), 57–62 (2023) https://doi.org/10.57041/ijeet.v2i1.916
  41. [41]Pan, J., Sotiriadis, P., Nenchev, V., Englberger, F.: Improving functional reliability of near-field monitoring for emergency braking in autonomous vehicles. In: 2025 IEEE 28th International Conference on Intelligent Transportation Systems (ITSC) (2025). https://doi.org/10.1109/ITSC60802.2025.11423625
  42. [42]Rusu, R.B., Cousins, S.: 3D is here: Point Cloud Library (PCL). In: Proc. IEEE International Conference on Robotics and Automation (ICRA), Shanghai, China, pp. 1–4 (2011). https://doi.org/10.1109/ICRA.2011.5980567
  43. [43]Kuhn, H.W.: The Hungarian method for the assignment problem. Naval Research Logistics Quarterly 2(1-2), 83–97 (1955) https://doi.org/10.1002/nav.3800020109

Paper details

vlnc:W_v3vjntwf
Version
v1 · latest
Published
Fresh — added to the Hub within the last 30 days
Added
License
CC BY 4.0
Usage

views · 0

downloads · 0

A view counts once per browser per day, a download each time the PDF is requested.

  • Computer Science

Not recorded for this paper

View statement
Looking for related papers…
Contents