RGB-D Object Tracking with Occlusion Detection
Yujun Xie, Yao Lu, Shuang Gu · 2019
Tracking objects in RGB-D images is a challenging task in computer vision, especially under occlusion. In this paper, we proposed an object tracking method based on the 3D point cloud. Firstly, we convert RGB-D images to point clouds. Secondly, features of point clouds are extracted by PointNet and finally integrated into the 3D object tracking algorithm for template matching across frames. A strategy of occlusion detection and target retrieval is applied to handle target missing under occlusion. For example, when the number of point clouds is decreasing abruptly, the occlusion may take place. Then a YOLOv3 detector is used to re-find this target. Our network is insensitive to appearance variation of object and robust to object tracking. The experimental results show that the proposed method achieves comparable results to state of the art on the Princeton RGB-D Tracking Benchmark.