Sequential Distance Dependent Chinese Restaurant Processes for Motion Segmentation of 3D Lidar Data

This paper proposes a novel object segmentation method for 3D Light Detection and Ranging (LIDAR) data that is particularly useful for the traffic scene analysis of self-driving vehicles. The novel method gains robustness against under-segmentation, i.e. the problem of assigning several objects to one segment, by jointly using geometrical features and motion field information to discriminate even spatially close objects in the data. The approach maps point cloud data to an occupancy grid representation and estimates the motion field using Kalman filter based tracking of grid cells. A non-parametric Bayesian clustering approach based on a sequential distance dependent Chinese Restaurant Process (s-ddCRP) utilizes this information in order to sample possible data segmentation hypotheses and decide on the most probable one. The computational efficiency of the approach is improved by exploiting the sequential nature of the problem and initializing the required Gibbs sampler using data from the previous time step and fixing found clusters of objects in super grid cells which can be tracked jointly. Experiments carried out on data obtained with a Velodyne HDL64 scanner in a real traffic scenario illustrate the performance of the approach.