Kalman Filter-Based Fusion of LiDAR and Camera Data in Bird's Eye View for Multi-Object Tracking in Autonomous...

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• Accurate multi-object tracking (MOT) is crucial for autonomous vehicles (AVs) but challenged by single-sensor limitations.
• Existing multi-modal MOT methods often depend on complex deep learning fusion.
• Robust tracking during occlusions remains a significant hurdle for AV perception.

Purpose of the Study:

• To introduce CLF-BEVSORT, a novel camera-LiDAR fusion model for 3D multi-object tracking (3DMOT) in autonomous driving.
• To enhance tracking robustness by fusing 2D camera and 3D LiDAR data in bird's eye view (BEV).
• To improve track recovery during short occlusions by leveraging LiDAR depth information.

Main Methods:

• Developed CLF-BEVSORT, a camera-LiDAR fusion model operating in bird's eye view (BEV).
• Integrated a novel association strategy incorporating structural similarity into the cost function for data fusion.
• Utilized the SORT tracking framework and leveraged LiDAR depth for robust track recovery.

Main Results:

• Achieved state-of-the-art HOTA score of 77.26% for Cars on the KITTI dataset, surpassing existing methods.
• Outperformed other approaches for Pedestrians with a HOTA score of 46.03%.
• Reduced identity switches (IDSW) by over 45% for cars compared to baseline methods.

Conclusions:

• CLF-BEVSORT demonstrates superior performance in 3D multi-object tracking for autonomous driving.
• The proposed fusion strategy effectively handles sensor data and improves tracking consistency.
• CLF-BEVSORT establishes a new benchmark for robust and accurate 3DMOT in dynamic environments.