SP-Dynamic-obstacle-detection-for-autonomous-gokarts🏎️

Project Overview

This repository documents my semester project at ETH Zurich's Institute for Dynamic Systems and Control (IDSC), focused on developing a robust obstacle detection system for autonomous racing go-karts. The project introduces a novel approach using a dynamic occupancy grid in a local frame, enhancing reliability through localization independence.

Note: The implementation code remains private within IDSC's repository.

Key Features

• Localization-Independent Operation: System operates in a local frame, reducing dependency on precise localization
• Real-time Processing: Achieves 20Hz update rate, matching LiDAR data acquisition frequency
• Dynamic Obstacle Tracking: Uses particle filter for tracking moving obstacles
• Motion Compensation: Implements Extended Kalman Filter for accurate state estimation
• Free Space Interpolation: Advanced algorithm for reducing grid sparsity while maintaining accuracy

Technical Implementation

1. Dynamic Occupancy Grid

• Discretized 2D representation of the environment
• Probabilistic framework for distinguishing between occupied, free, and unknown spaces
• Integration of LiDAR measurements with interpolation for comprehensive spatial mapping

2. Motion Compensation

• Extended Kalman Filter (EKF) for state estimation
• Kinematic bicycle model for go-kart dynamics
• Integration of multiple sensor inputs:
  • Wheel encoders
  • Steering sensor
  • IMU
  • LiDAR with ICP

3. Particle Filter for Dynamic Obstacles

• State representation including position and velocity (x, y, vx, vy)
• Real-time tracking and prediction of obstacle movements
• Adaptive resampling for maintaining particle diversity

Results and Performance

Achievements

• Successfully implemented real-time processing at 20Hz
• Effective obstacle detection in static environments
• Reliable tracking of dynamic obstacles under moderate maneuvers
• Robust free space interpolation reducing grid sparsity

Current Limitations

• Performance degradation during high-yaw maneuvers
• Computational constraints on CPU limiting particle count
• Assumptions about pitch and roll occasionally affecting accuracy
• Integration with planning systems pending

Visualization showing the limitation of particle filter during high-yaw maneuvers

Future Work

Planned Improvements

1. GPU Implementation

   • Parallelization of particle filter computations
   • Support for increased particle count and grid resolution

2. Motion Planning Integration

   • Development of local frame planner
   • Interface with existing MPC trajectory planning

3. System Enhancements

   • Quantitative analysis and parameter optimization
   • Complex motion models for better obstacle prediction
   • Sensor fusion with camera data

Academic Context

This project was conducted as a semester project at:

• Institution: ETH Zurich
• Department: Institute for Dynamic Systems and Control (IDSC)
• Supervision: Dr. Maurilio Di Cicco, Prof. Dr. Emilio Frazzoli
• Date: February 2024

References

1. AMZ Driverless: The Full Autonomous Racing System
2. Random Finite Set Approach for Dynamic Occupancy Grid Maps with Real-time Application
3. Extended Kalman Filter Documentation

Note: Implementation details and code are maintained privately within IDSC's repositories.