AHMOT: Adaptive Kalman Filtering and Hierarchical Data Association for 3-D Multiobject Tracking in IoT-Enabled Autonomous Vehicles

Accurate and robust 3-D multiobject tracking (3-D MOT) is necessary to ensure the safe operation of autonomous vehicles in complex traffic environments. However, existing 3-D MOT methods frequently face challenges stemming from dynamic environmental noise, real-time processing demands, and the inherent limitations of individual sensors. To address these issues, an enhanced 3-D MOT framework that integrates adaptive Kalman filtering (AKF) with a novel 3-D hierarchical data association model is presented in this article. In contrast to traditional Kalman filters, the proposed AKF automatically adjusts filter parameters in real time to account for environmental fluctuations and model uncertainties, thereby improving both tracking accuracy and system robustness. Additionally, the 3-D hierarchical data association model differentiates itself from existing methods by comprehensively incorporating the 3-D intersection-over-union (IoU), appearance, and distance features of targets while employing a cascade strategy that enhances real-time performance. By leveraging an Internet of Things (IoT) framework, the proposed system seamlessly fuses data from LiDAR and camera sensors, capitalizing on the scalability and real-time processing of IoT to overcome the constraints of individual sensors. On the KITTI and nuScenes datasets, our method achieves a 6.8% improvement in the tracking accuracy metric sAMOTA compared with the state-of-the-art methods and attains a superior HOTA score of 51.63%, outperforming all the compared methods. Moreover, our method operates at a real-time processing rate of 92 FPS. Collectively, these advancements enable more accurate and stable multiobject tracking in different traffic settings, ultimately enhancing the safety and reliability of autonomous driving systems in real-world scenarios.