Related Experiment Video
Updated: Jul 13, 2025
14:55
Evaluating the Effect of Roadside Parking on a Dual-Direction Urban Street
Published on: January 20, 2023
3.4K
Vehicle Detection and Tracking with Roadside LiDAR Using Improved ResNet18 and the Hungarian Algorithm
Ciyun Lin1,2, Ganghao Sun1, Dayong Wu3
1Department of Traffic Information and Control Engineering, Jilin University No. 5988, Renmin Street, Changchun 130022, China.
Sensors (Basel, Switzerland)
|October 14, 2023
Summary
A new vehicle tracking algorithm using roadside LiDAR infrastructure achieves 100 ms latency for autonomous driving. This system enhances detection accuracy and tracking performance, crucial for Level 5 autonomy and real-time transportation applications.
Area of Science:
- Computer Vision
- Robotics
- Transportation Engineering
Background:
- Achieving Level 5 autonomous driving necessitates vehicle-infrastructure collaboration, requiring low-latency, high-speed data processing.
- Current systems face challenges in maintaining detection accuracy and reducing latency for real-time applications.
Purpose of the Study:
- To introduce a vehicle tracking algorithm utilizing roadside LiDAR infrastructure to achieve low latency (100 ms) without compromising detection accuracy.
- To enhance vehicle detection and tracking for improved performance in autonomous driving scenarios.
Main Methods:
- Developed a vehicle detection architecture using ResNet18, optimizing Bird's-Eye View (BEV) mapping and loss functions for full frame rate detection.
- Proposed a three-stage vehicle tracking algorithm enhancing the Hungarian algorithm with time-space logicality and trajectory similarity to handle occlusions.
- Tested the framework on the KITTI dataset and a MATLAB/Simulink simulation dataset.
Main Results:
- Achieved F1-scores of 96.97% (KITTI) and 98.58% (MATLAB/Simulink) for vehicle detection.
- Obtained MOTA scores of 88.12% (KITTI) and 90.56% (MATLAB/Simulink), and ID-F1 scores of 95.16% (KITTI) and 96.43% (MATLAB/Simulink) for vehicle tracking.
- Demonstrated significant improvements in calculation speed compared to traditional methods.
Conclusions:
- The proposed LiDAR-based vehicle tracking algorithm effectively reduces latency to 100 ms while maintaining high detection and tracking accuracy.
- The enhanced tracking algorithm outperforms traditional methods, particularly in handling occlusions and improving computational speed.
- This framework is a promising solution for enabling real-time performance required for Level 5 autonomous driving.
