Enhancing Real-Time Multi-Object Tracking with Pillar-Based Models: Integration of Class Information and BiFPN Neck

In recent years, center-based tracking has been the primary approach for multi-object tracking (MOT) due to its speed and accuracy. However, 3D MOT operates in much more complex environments than 2D, facing challenges such as occlusion, object deformation, and real-time processing difficulties. Addressing these challenges is critical for real-time applications such as autonomous driving, robotic vision, and surveillance systems, where accurate and fast tracking directly impacts safety and efficiency. For instance, in autonomous vehicles, poor object tracking performance can lead to accidents or incorrect route planning. Therefore, ensuring accurate object detection and tracking performance in real time is essential in these fields. To address this issue, this study proposes Enhancing Real-Time Multi-Object Tracking with Pillar-Based Models, which applies BiFPN instead of the conventional FPN to enhance feature extraction and integrates object class information into the velocity estimation head for more precise tracking. Evaluation on the nuScenes dataset demonstrates that the proposed method significantly improves 3D MOT performance, resulting in a 5.8% increase in AMOTA and a reduction of IDS by 133.