Active Perception: Maneuvers AVs Use to Reduce Occlusion Risk

/ General / By

When parts of the scene are hidden, perception alone is not enough: the vehicle must change its configuration or motion to collect more informative sensor data. Below are the common, practical active-perception maneuvers used by production and research AV stacks, and how to weigh their benefits and costs.

Creep forward (controlled inching)

Brief, low-speed forward motion from a stopped or slow state to reveal hidden areas (e.g., past a parked van or around a corner).

When to use: narrow sight-lines, low cross-traffic risk, or when sensors need a short baseline to detect pedestrians.

How to implement: limit speed (typically 0.5–5 km/h), short distance steps with frequent re-evaluation, and a hard cutoff for full stop if new threats appear.

Lane offset and lateral repositioning

Small lateral shifts within the lane or to an adjacent safe lane to change the camera/lidar vantage point without entering opposing traffic.

When to use: occlusion caused by roadside objects (parked vehicles, vegetation) and when a lateral view exposes hidden sidewalks or crosswalks.

How to implement: verify adjacent lane is clear, signal and execute a controlled offset, maintain safe clearance from obstacles and curb, and monitor for oncoming cyclists or vehicles.

Speed modulation and early deceleration

Reducing speed earlier than strictly required both increases reaction time and gives sensors more frames to accumulate evidence.

When to use: high-uncertainty zones like driveways, blind crosswalks, or busy urban corners.

How to implement: blend comfort-aware deceleration profiles with conservative minimum speeds; combine with driver-courtesy signaling (hazard lights) where appropriate.

Temporary stop and observational hold

Full stop with active sensing (rotating lidar heads, attention-focused cameras) to wait for a safe gap or clearer observation.

When to use: very high occlusion with likely vulnerable road users (children, bicycles) or when motion would create unacceptable risk.

How to implement: define maximum dwell time, escalate to alternative actions (reroute or request remote operator) if no resolution, and avoid blocking critical lanes.

Sensor-centric maneuvers (pan, tilt, active sensors)

Physically reorienting sensors (camera pan/tilt, steerable lidar) or using vehicle motion to create a parallax baseline for depth estimation.

When to use: when mechanical reorientation or slight motion produces substantial new information without altering traffic position.

How to implement: coordinate sensor actuation with motion planner and perception stack to avoid blind intervals; prefer small, frequent adjustments over large, disruptive motions.

Cooperative strategies and information augmentation

Use external data to reduce the need for risky motion: V2X messages, infrastructure cameras, maps with occlusion priors, or cueing from other road users (e.g., lead vehicle braking).

When to use: where infrastructure or other agents are available and trustworthy; as a complement to motion-based strategies.

How to implement: validate external inputs, attach confidence weights, and fall back to conservative maneuvers if data is missing or inconsistent.

Trade-offs: safety, comfort, and traffic flow

Safety-first policies favor longer stops and lower speeds but can cause traffic delays and abrupt interactions. Comfort-oriented policies limit jerky creep moves and frequent stops but accept higher uncertainty. Balanced planners use risk metrics (probability of emergence × severity) to pick the least-cost action that keeps collision risk below threshold while minimizing delay and discomfort.

Practical planner rules-of-thumb

  • Prefer viewpoint changes that stay within legal lanes (lane offset over crossing a double yellow).
  • Use incremental steps: creep small distances with frequent reassessment rather than one long uncontrolled move.
  • Combine signals: slow + lateral offset + sensor reorientation often gives the best information per unit risk.
  • Cap dwell and creep times to avoid blocking intersections or creating secondary hazards; escalate to alternative routes if unresolved.
  • Blend priors with sensing: map-based occlusion likelihoods should modulate how aggressive viewpoint maneuvers are.

Active perception is a set of low-speed, information-gathering maneuvers chosen by the behavior planner to reduce hidden risk while respecting traffic rules and comfort. Implemented conservatively—short creeps, safe lateral offsets, coordinated sensor actuation, and strict reassessment—these maneuvers let an AV behave like a cautious, competent driver when sightlines are limited.

Sources

Erhalten Sie alle aktuellen Neuigkeiten und Informationen direkt in Ihrem Posteingang.

e Deutsch