Using Surround-Vehicle Maneuvers as Indirect Perception Cues

When parts of the scene are occluded, the motion of visible road users becomes a vital information source. Rather than treating other vehicles, cyclists, and pedestrians as independent obstacles, modern autonomous systems extract indirect cues from their maneuvers to reduce uncertainty about what may be hidden around corners or behind parked vehicles.

Key motion cues and what they imply

Deceleration / brake lights: A sudden slow-down or illuminated brake lamps on a lead vehicle near a crosswalk or driveway raises the probability of a pedestrian or stopped vehicle ahead. For planners this typically increases predicted risk in the occluded region and causes earlier speed reduction or a higher-probability short-horizon stop.

Steering yaw and lateral shifts: Lateral motion toward an exit, shoulder, or curb often signals an intent to turn, park, or avoid an obstacle. If that lateral motion occurs near an occlusion (e.g., at a corner), the AV should increase the likelihood of a hidden actor emerging along the intended path.

Turn signals and indicators: Blinkers provide explicit intent cues; timely use reduces uncertainty about lane changes or turns. Systems should fuse signal state with motion to avoid over-relying on signals (which may be missing or mistimed).

Stopping/hesitation patterns: Intermittent speed oscillation, chesting, or brief stops by road users (including other drivers) near crosswalks often indicate scanning or yielding for a vulnerable road user. These patterns are a weak but consistent predictor of pedestrian emergence.

Inter-actor interactions: A following vehicle braking in response to a distant vehicle can be strong indirect evidence that something unseen is causing the lead vehicle to slow; similarly, vehicles yielding to an unseen gap increase the probability of crossing traffic in the occluded lane.

How to represent motion cues in perception and prediction

Probabilistic intent priors: Convert observed maneuvers into updated priors over hidden-state hypotheses (e.g., pedestrian present vs. absent) using Bayesian updates or particle filters. Weight priors by context (location type, speed limit, map semantics).

Behavior-conditioned occupancy: Augment occupancy grids or bird’s-eye-view (BEV) maps with behavior-conditioned probability fields: e.g., if a vehicle ahead brakes near a crosswalk, raise occupancy probability in the adjacent occluded sidewalk cells for the next 2–5 s.

Multi-agent interaction models: Use social or graph-based predictors that model how one agent’s motion changes another’s distribution—this captures yielding, gap acceptance, and cooperative evasive responses that hint at hidden hazards.

Short-horizon detectors for “indirect events”: Train classifiers to detect signatures (rapid brake, emergent lateral jerk, clustered decelerations) that historically precede hidden-object encounters; trigger conservative fallback behaviors when high-confidence.

Integration into planning and safety stacks

Risk-aware planning: Translate increased hidden-object probability into explicit risk cost terms (speed reduction, larger safety buffer, or active information-gathering maneuvers such as creeping forward or adjusting viewpoint).

Active perception maneuvers: When motion cues raise uncertainty, prefer small exploratory actions (e.g., inching forward, slight lateral offset) that improve line-of-sight while preserving an easy-stop trajectory.

Conservative fallback policies: Define graded responses tied to cue strength: soft slowdowns for weak cues, full stops and alert states for strong, corroborated cues (multiple actors braking, visual confirmation of brake lights plus lateral avoidance).

Practical considerations and failure modes

False positives from noisy or aggressive driving: Harsh braking or swerving for speed control (not for hazards) can mislead belief updates—use context checks (location semantics, speed, traffic rules) and temporal smoothing to reduce spurious responses.

Missing or incorrect signals: Do not treat turn signals or brake lights as authoritative; combine them with motion and map priors. Account for occluded or noncompliant actors (e.g., cyclists without signals).

Latency and sensing limits: Indirect cues are only useful if detected and interpreted quickly; ensure low-latency detection and align prediction horizons so planners can act before the hidden actor appears.

Example pipeline (practical recipe)

1) Detect maneuvers on visible actors (braking, steering, signaling) with short temporal windows (0.5–2 s).

2) Convert detections into likelihood updates for hidden-object hypotheses using context-weighted Bayesian rules (map semantics, speed, time-of-day).

3) Update behavior-conditioned occupancy/B EV and pass the higher-risk map to a risk-aware planner that applies a graded slowdown or viewpoint-improving maneuver.

4) If multiple corroborating cues accumulate within the planning horizon, escalate to full stop and request visual confirmation before proceeding.

Summary

Motion cues from visible actors provide actionable, often early evidence of hidden hazards. Reliable use requires fusing these cues with map priors, contextual filters, and conservative planning rules that favor active perception and graded fallbacks. Properly implemented, surround-vehicle maneuvers become an inexpensive, robust extension of perception for safer behavior in occluded scenarios.

Sources

العربية