BEVSORT: Bird Eye View LiDAR Multi Object Tracking

Abstract

Multi-object tracking (MOT) is an essential task for robotic solutions, requiring 3D information about objects. Cameras can provide some 3D information, but LiDAR sensors can provide more accurate and reliable 3D data. Recent deep learning tracking techniques have achieved top performance on public datasets; however, they are too complex for edge devices. Filter-based tracking techniques are mature and still used in embedded robotics and automotive applications, but they need to be adapted to handle more complex scenarios. This work adopts Kalman filter-based trackers, proposing a new SORT variant for tracking objects in point cloud addressing the limitations of linear motion assumptions and varying the tracked states to adapt objects in bird eye view perspective.