Prediction Module

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

Overview

Predicts future trajectories for tracked objects by:

1. Retreiving object type from Preception (vehicle/pedestrian/cyclist)
2. Querying HD map for current and possible future lanelets
3. Generating multiple trajectory hypotheses (different intents/paths)
4. Assigning probabilities to each hypothesis

Current Status: Skeleton with placeholder implementations. Runs standalone without map services.

ROS Interface

Subscribed Topics

Published Topics

Services Used

Map queries to /world_modeling/lanelet/query/* (placeholders until map services available)

Architecture

Component-based design for parallel team development:

• prediction_node: Orchestrates pipeline, handles ROS communication
• trajectory_predictor: Generates hypotheses (person 1: pedestrian, person 2: vehicle, person 3: cyclist)
• motion_models: Physics-based propagation (bicycle model, constant velocity)
• intent_classifier: Assigns probabilities (shared by all)
• map_interface: HD map queries (currently placeholders)

Each component has single responsibility and can be tested independently.

Quick Start

# Build
colcon build --packages-select prediction

# Run
ros2 launch prediction prediction.launch.py

Team Tasks

Person 1 - Pedestrian Prediction:

• File: src/trajectory_predictor.cpp → generatePedestrianHypotheses()
• File: src/motion_models.cpp → Add noise to ConstantVelocityModel
• Use constant velocity with Gaussian noise, goal-directed behavior at crosswalks

Person 2 - Vehicle Prediction:

• File: src/trajectory_predictor.cpp → generateVehicleHypotheses()
• File: src/motion_models.cpp → BicycleModel::generateTrajectory()
• Implement bicycle kinematics with path following (pure pursuit or Stanley controller)

Person 3 - Cyclist Prediction:

• File: src/trajectory_predictor.cpp → generateCyclistHypotheses()
• Research cyclist behavior, implement hybrid model
• Use pedestrian model at crosswalks, vehicle model on roads
• Critical: Coordinate output format with Person 1 & 2

Output Format (all must match):

struct TrajectoryHypothesis {
  std::vector<geometry_msgs::msg::Pose> waypoints;
  std::vector<double> timestamps;
  Intent intent;
  double probability;  // Set by classifier
};

Configuration

Parameters in config/params.yaml:

• prediction_horizon: 5.0 seconds
• prediction_time_step: 0.1 seconds
• Vehicle/pedestrian/cyclist specific parameters

Team Assignments

• Girish: Pedestrian prediction system (constant velocity model)
• John: Vehicle prediction system (bicycle kinematics)
• Aruhant: Cyclist prediction system (hybrid model)

See inline code comments marked with names for specific tasks.

Current Limitations

File truncated at 100 lines see the full file