Designing Contextual Filters to Reduce False Positives from Motion Cues

/ General / By

When surrounding actors move, their maneuvers are useful indirect cues but also a common source of false positives (e.g., a follower braking for scenery, a signaling lane change that is aborted). Contextual filters are lightweight, sensor-agnostic layers that screen motion cues before they influence planning. This article outlines pragmatic filter types, classifier approaches, and conservative fallback rules that integrate cleanly with map priors and active-perception policies.

1) Objectives and design constraints

Objectives: reduce spurious hazard alarms from benign maneuvers, maintain early detection of real threats, and keep computational cost and latency low.

Constraints: should be sensor-agnostic (work from object tracks), respect temporal continuity, be explainable for safety cases, and degrade gracefully under uncertainty.

2) Filter categories

Spatial-context filters: use lane geometry, crosswalks, curb presence, and parking layout (from HD map or local map estimates) to downweight improbable hazard interpretations — e.g., a vehicle braking entirely within its lane near a stop sign is less likely to imply a hidden crossing than one braking near an occluding parked van at an intersection.

Kinematic-consistency filters: compare observed accelerations, yaw rates, and lateral offsets against typical envelopes for the actor class (car, bike, pedestrian). Short, low-magnitude decelerations or micro-lane corrections are often benign; sustained decelerations that violate normal envelope raise the hazard score.

Inter-actor relational filters: consider neighbors’ responses. If a following vehicle brakes but the lead vehicle shows no upstream deceleration (within sensor noise and latency), treat the follower’s brake as less informative. Conversely, correlated braking across multiple independent actors increases hazard probability.

Semantic-signal filters: leverage turn signals, brake lights, and pedestrian gestures when available (from vision or CAN). A signaled lane change that completes within expected time reduces crossing-threat score; an un-signaled sudden lateral move increases it.

Temporal-persistence filters: require minimal temporal support before escalating a hazard. Use graded thresholds (e.g., 1 frame = ignore, 2–4 frames = monitor, ≥5 frames = escalate) tuned per actor class and speed.

3) Lightweight heuristic scoring pipeline

Build a per-actor hazard score by combining filter outputs with simple weights:

1) base_score from raw motion cue (e.g., follower braking magnitude),

2) multiply by spatial_context_factor (0.0–2.0),

3) multiply by kinematic_consistency_factor (0.0–2.0),

4) multiply by relational_factor (0.0–2.0),

5) apply persistence_boost if temporal support exceeds thresholds.

Map the final score into three bands: monitor (low), active-perception (medium), and immediate-planning-adjustment (high).

4) ML classifiers and hybrid models

When to use ML: for complex relational patterns or when labeled data exists for distinguishing benign vs hazardous maneuvers.

Model choices: compact temporal models (1D CNN, GRU) on per-actor time series; graph models for inter-actor relations; and tree ensembles for explainability where features are handcrafted filter outputs.

Training targets and labels: binary hazard vs benign, and ordinal severity (monitor/active/immediate). Use data augmentation for occlusion scenarios and balance by actor class and environment.

Explainability: prefer models that output feature attributions (SHAP, attention scores) or combine ML outputs with the heuristic pipeline so final decisions remain inspectable.

5) Calibration, tuning, and safety margins

Calibrate thresholds per operational design domain (ODD): urban dense, suburban, highway. Enforce conservative bias: when uncertain, escalate to active-perception (inch forward, lateral offset) rather than immediate aggressive avoidance. Log filter activations and errors for continuous improvement.

6) Integration with planning and active perception

Connect hazard bands to graded planner actions: monitor = maintain speed and continue sensing; active-perception = execute safe exploratory maneuver (small forward step, lateral peek) while preserving an easy-stop; immediate = slow or stop with clear safety margin. Always combine with map priors and collision-avoidance invariants.

7) Failure modes and mitigations

Over-suppression: filters that are too aggressive can mask real hazards — mitigate by requiring at least one independent high-sensitivity cue (e.g., multiple actors or map conflict) to unblock suppression.

Sensor bias or missing signals: design filters to operate on tracks rather than raw sensors and fall back to persistence and relational cues if semantic signals are absent.

8) Operational checklist for deployment

– Implement map-backed spatial filters and kinematic envelopes for each actor class.

– Add temporal persistence thresholds and inter-actor correlation checks.

– Train a lightweight temporal classifier or tree ensemble to resolve ambiguous cases; retain explainability hooks.

– Define graded planner interfaces and conservative default behaviors under uncertainty.

– Monitor performance and iterate thresholds per ODD using logged cases.

Contextual filters are a pragmatic middle layer that substantially reduces false positives from motion cues while preserving early hazard detection. When designed sensor-agnostically, with clear failure modes and conservative fallbacks, they improve robustness for occlusion-heavy driving scenarios.

Sources

Λάβετε όλα τα τελευταία νέα και πληροφορίες στα εισερχόμενά σας.

Ελληνικά