Prediction Module

Multi-modal trajectory prediction for vehicles, pedestrians, and cyclists using physics-based motion models with lanelet-aware intent inference.

Overview

Predicts future trajectories for tracked objects by:

1. Retrieving object type from Perception (vehicle/pedestrian/cyclist)
2. Querying HD map for reachable lanelets around the object
3. Generating multiple trajectory hypotheses using motion models
4. Assigning probabilities to each hypothesis with temporal smoothing

Current Status: Fully implemented with lanelet-aware prediction, per-vehicle caching, and async service queries.

ROS Interface

Subscribed Topics

Published Topics

Services Used

Architecture

Modular component design:

• PredictionNode: Lifecycle management, ROS communication, temporal smoothing
  • Subscribes to detections, ego pose, ego-relative lanelets
  • Manages async per-vehicle lanelet service requests
  • Applies confidence smoothing to reduce frame-to-frame flicker
  • Publishes world objects with trajectory hypotheses
• TrajectoryPredictor: Hypothesis generation with lanelet awareness
  • generateHypotheses(): Routes to type-specific generators
  • generateLaneletVehicleHypotheses(): Path-following hypotheses (left/right/straight)
  • generateGeometricVehicleHypotheses(): Fallback when no lanelet data
  • generatePedestrianHypotheses(): Constant velocity with intent variation
  • generateCyclistHypotheses(): Hybrid vehicle/pedestrian behavior
  • Per-vehicle lanelet caching with invalidation distance
  • Speed estimation from position history
• MotionModels: Physics-based trajectory propagation
  • BicycleModel: Kinematic bicycle model for vehicle trajectories
  • ConstantVelocityModel: Simple velocity propagation for pedestrians
• IntentClassifier: Probability assignment to hypotheses
  • Geometric scoring (heading alignment, lanelet match quality)
  • Maneuver priors and inertia
  • Trajectory smoothness penalties

Each component has single responsibility and clear interfaces.

Quick Start

# Build prediction module and dependencies
colcon build --packages-select prediction world_model

# Run prediction node with world model
ros2 launch prediction prediction.launch.py

Key Features

Lanelet-Aware Prediction

• Queries reachable lanelets around detected vehicles via get_lanelet_ahead service
• Per-vehicle caching prevents redundant service requests within 5m movement threshold
• Falls back to geometric prediction when lanelet data unavailable

Temporal Smoothing

• Confidence smoothing (α-filter) reduces hypothesis flickering between frames
• Matches hypotheses by intent and endpoint location (6m threshold)
• Timeout removes stale object state after 5 seconds

Async Service Queries

• Non-blocking per-vehicle lanelet queries using ROS2 async service clients
• Limits concurrent requests (max 8 pending) to prevent service overload
• Maintains per-vehicle cache keyed by detection ID

Speed Estimation

• Tracks position history per object for velocity estimation
• Falls back to bounding box length heuristic when history unavailable
• Used to parameterize motion models

Configuration

File truncated at 100 lines see the full file