Friday 31 January 2025
Self-driving cars are getting better at navigating complex environments, thanks in part to advancements in multi-object tracking and simultaneous localization and mapping (SLAM). Researchers at Shanghai Jiao Tong University have developed a new system that tightly couples LiDAR SLAM and confidence-guided implicit data association for real-time multi-object tracking.
The team’s Conf SLAMMOT method uses a combination of LiDAR odometry, object detection, and graph optimization to track multiple objects in 3D space. The system is designed to handle challenging scenarios where objects are occluded or distant, making it more robust than previous methods.
Conf SLAMMOT works by first using LiDAR odometry to estimate the ego-vehicle’s pose and map the environment. Then, object detection algorithms like PV-RCNN are used to identify potential targets. The system then uses a confidence-guided implicit data association approach to match detected objects with their predicted states.
The team tested Conf SLAMMOT on the KITTI tracking dataset, which is commonly used in autonomous vehicle research. They compared their method to other state-of-the-art methods like SECOND, PointRCNN, and LIO-SEGMOT. The results showed that Conf SLAMMOT achieved competitive accuracy and performance in terms of ego-pose estimation and object state estimation.
One of the key advantages of Conf SLAMMOT is its ability to recover from missed detections caused by occlusion or distance. In these scenarios, other methods may struggle to maintain accurate tracking. However, Conf SLAMMOT’s confidence-guided implicit data association approach allows it to adaptively adjust the implicit object association range, reducing the impact of missed detections.
The team also tested Conf SLAMMOT on a variety of sequences from the KITTI dataset, including scenarios with multiple objects and complex motion patterns. The results showed that Conf SLAMMOT was able to accurately track objects in these challenging environments.
Overall, Conf SLAMMOT represents an important advancement in multi-object tracking and SLAM for autonomous vehicles. Its ability to handle challenging scenarios and recover from missed detections makes it a valuable tool for developers working on self-driving car technology.
Cite this article: “Confident Object Tracking in Autonomous Vehicles: A Novel Method Combining LiDAR SLAM and Implicit Data Association”, The Science Archive, 2025.
Lidar, Slam, Multi-Object Tracking, Autonomous Vehicles, Conf Slammot, Kitti Dataset, Object Detection, Graph Optimization, Confidence-Guided Implicit Data Association, Ego-Pose Estimation, State Estimation
Reference: Susu Fang, Hao Li, “LiDAR SLAMMOT based on Confidence-guided Data Association” (2024).
