The precise perception of moving objects in a dynamic environment is an essential task for autonomous vehicles. Existing object detection methods using point clouds are difficult to distinguish the moving and static objects. Inspired by the excellent performance of the optical flow method in moving objects detection in the image domain, we adopt optical flow method for point cloud data and propose a new moving object detection method using point clouds. The proposed new method contains semantic constraints and a scene flow module. In detail, we first fuse semantic information to judge static objects, such as ground and buildings, and exclude these points of static objects; then we apply the scene flow module to estimate the motion vector field of two consecutive point clouds and predict moving or static labels to each point. At the end, exhaustive experiments with a public SemanticKITTI dataset are conducted for validation and evaluation. The results demonstrate the competitive performance of the proposed method comparing to the existing state-of-the-art methods.

moving object detection;dynamic environment;point cloud;semantic;scene flow