A Novel Multiobject Tracking Framework Based on Multisensor Data Fusion for Autonomous Driving in Adverse Weather Environments

• Published in: IEEE sensors journal Vol. 25; no. 9; pp. 16068 - 16079
• Main Authors: Han, Liuyang, Wang, Jun, Li, Chen, Tao, Fazhan, Fu, Zhumu
• Format: Journal Article
• Language: English
• Published: New York IEEE 01.05.2025; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Adverse weather environments; Algorithms; autonomous driving (AD); Cameras; Data integration; Datasets; Laser radar; Meteorology; multiobject tracking (MOT); Multiple target tracking; multisensor data fusion; Multisensor fusion; Noise; Point cloud compression; progressive association; Real time; Sensors; Target tracking; Three-dimensional displays; Trajectory; Weather
• ISSN: 1530-437X; 1558-1748
• DOI: 10.1109/JSEN.2025.3550506

Multiobject tracking (MOT) is a critical task for autonomous driving (AD), aiming to localize and maintain consistent identities of targets in camera or LiDAR data. However, adverse weather conditions (low-light, rain, fog, and sandstorms) severely degrade sensor data quality, leading to deteriorated detection accuracy (DetA) and unstable tracking performance in existing methods. This study proposes a robust multisensor fusion framework for adverse weather situations to address these challenges. Specifically, a physics-based adverse weather simulation algorithm is introduced to generate the Adverse dataset. Furthermore, a geometrically verified late-fusion detection method is proposed to resolve conflicts between 2-D and 3-D detection results, enhancing detection robustness. In addition, a spatial-appearance progressive association strategy is developed for stable tracking. Experimental results demonstrate the method's superior performance across multiple datasets, particularly under adverse weather conditions. On the KITTI dataset, the method achieves state-of-the-art results in HOTA (80.13%), MOTP (87.69%), and localization accuracy (LocA) (88.84%), outperforming all baseline methods. In addition, it meets the real-time performance requirements for AD deployments.