3D-Multi-Hypothesis Tracker for Multi-Object Tracking Applications

Rapid advances are being made in the automotive electronic sensor technologies, enabling 3D sensing and processing capabilities for autonomous vehicles. Multi-hypothesis tracking (MHT) provides a systematic solution for data association in multiple object tracking scenarios from traditional 2D video cameras. However, MHT incurs large memory and computational costs and cannot be directly applied in 3D point cloud data. We introduce 3D-based (3D-MHT) method. We have incorporated 3D bounding box and detection information into score updating and measurement procedure. We address the limitation of exponential growth of hypotheses in the traditional MHT method by incorporating the k-best scores and N-scan method during the data association to control the number of hypotheses. Our 3D-MHT approach achieved a 2.72% higher accuracy and 7.92% higher precision compared to the baseline method on KITTI, a large 3D dataset.