3D Multi-Object Tracking Based on Point Cloud Scene Flow Estimation for Autonomous Driving

We design the scene flow estimation network with a new cylindrical coverage aware sampling strategy to predict point-wise motion flow, and density attentive mechanism is integrated. Then, we propose 3D multi-object tracking method  based on the scene flow estimation, and an adaptive volume Intersection over Union cost is designed. We complete  the training and validation of the proposed 3D multi-object tracking method  based on the scene flow estimation using 3D point cloud with KITTI dataset.