An Automatic Background Filtering Method for Detection of Road Users in Heavy Traffics Using Roadside 3-D LiDAR Sensors With Noises

Background filtering for roadside Light Detecting and Ranging (LiDAR) data could benefit detection of road users with high frequency and accuracy. But existing methods are unsatisfactory under traffic congestions and neglect noises of LiDAR sensors. This paper attempts to develop background identification and filtering methods for heavy traffics and LiDAR noises. In particular, a 3-D space-time diagram of the LiDAR’s detection space over frames is divided into plenty of subspace-frames, where one subspace-frame refers to one subspace from the isometric division of detection space in one frame. For subspace-frames at the same position, an unsupervised clustering method with unknown number of clusters is applied for classifying subspace-frames into different groups and the background group is identified based on background characteristics. The proposed clustering method is a
${t}$
-distribution mixture model with constraints, where weights follow a Dirichlet distribution so that some weights have the possibility to be 0. A Bayesian estimation framework is developed to carry out the clustering. Then each background subspace-frame is divided into cubes with equal lengths. By aggregating points over frames in the same cube, the cube with aggregated points is regarded as background cube. For background filtering, points in background cubes are background points and excluded. Finally, examined by data from a Velodyne VLP-32 LiDAR sensor, the proposed approach has better filtering results no matter under free-flow conditions or congestions. Using background filtering results for road users’ detection is introduced as well.