Object Tracking With 3D LIDAR Via Sparse Learning

As an active sensor for environment sensing, Lidar sensor has experienced rapid development since 1990's. Lidar emits laser rays to percept the depth information from external circumstances. It has the advantage of insensitiveness to illumination conditions and high reliability. Therefore Lidar is widely used in mobile robot navigation task.Moving object tracking is a fundamental task for autonomous vehicles operating under urban areas. This paper began its research on robust object tracking utilizing 3D point cloud data acquired by Lidar. Sparse learning has achieved a lot of successful experience in visual object tracking research, while still far from practical application dues to the limitation of visual sensors. Inspired by this, we exploit sparse learning on Lidar object tracking and developed a new tracking algorithm. Our algorithm based on particle filter tracking framework, and interprets tracking process as a multi-task multi-cue sparse learning process. We utilized a multi-cue object representation method to build an over-complete template dictionary, and then represent each target candidate represented with a sparse linear combination of the templates. The tracking problem is formulated as finding the candidate with the lowest reconstruction error. To improve the robustness of the algorithm, the depth information is further enhanced by a specifically designed object depth estimation and occlusion detection mechanism, which can effectively eliminate background noise and prevent the occlusion from updated into target templates. Massive experiments show promising object tracking performance with our method, especially when handling complex situations such as occlusion and posture change.