Detection of Abnormal Motion by Estimating Scene Flows of Point Clouds for Autonomous Driving

Detection of abnormal motion in the environment is one of the most challenging and, at the same time, an important field of research in autonomous driving. Detecting an abnormal motion of objects using video sequences has been widely studied in the last decade. This paper proposes a method to detect abnormal motion by using sequences of LiDARs' point-clouds. Estimation of scene flows enables to compute the motion field in 3D space. Nearest point loss and Cosine-Cycle Consistency loss are proposed to estimate the scene flows of point-clouds in an unsupervised way. The direction of motion of the nearest object is represented in a 3D grid-like structure called VoxCF. VoxCF is used as an input for the extraction of dynamic features that encode the direction of objects' motion. Dynamic features are used to make decisions about the abnormal motion by making predictions from the Long short-term Memory (LSTM) network. The proposed method has been validated by performing real experiments with AV, and the obtained results show the precise detection of abnormal motion.