E3-MOT: An Extended End-to-End Multiple Object Tracking Framework With Camera-LiDAR Fusion

3-D multiobject tracking (MOT) is essential for providing stable and reliable motion states of surrounding obstacles in autonomous driving. Existing methods primarily focus on motion-based and appearance similarity matching approaches. However, the nature of post-processing limits the exploitation of multimodal perception data, hindering the effectiveness of these methods. In this work, we propose E3-MOT, an extended end-to-end multimodal tracking framework that integrates camera and LiDAR information within a shared bird’s-eye view (BEV) representation. A two-stage mechanism is designed, where the first stage performs endto- end joint detection and tracking (JDT). The track query is then introduced to represent the tracked instances, which are transferred and updated across consecutive frames, enabling iterative prediction throughout the tracking process. In the second stage, we design a motion-based tracking filter, which enhances the robustness through the stage-2 association between unmatched detections and trajectories on the image plane to address the long-tail distribution challenge. Extensive experiments on the nuScenes dataset demonstrate the effectiveness of the proposed method. E3-MOT achieves 67.4% average multiobject tracking accuracy (AMOTA), and under sensor-failure conditions, it still maintains 62.5% AMOTA, outperforming multiple representative baselines. Real-world tests on a UGV platform further validate the practicality and robustness of the framework.