Lidar-based vehicle detection for dynamic SLAM applications

Detection of dynamic objects plays an important role in dynamic Simultaneous Localization and Mapping (SLAM). In this study, a vehicle detection algorithm based on Lidar point cloud data is proposed. Firstly, the initial model of a vehicle is established based on distribution of point clouds, and the confidence level of the initial model is calculated. According to the maximum confidence level of the initial model, a vehicle model can be (1) represented by the initial model if the fixed threshold of the confidence level is reached, or (2) built based on bounding lines that are formed by residual points after removal of sparse points, if the fixed threshold of the confidence level fails to be reached. This method applies not only to detection of vehicles with a distinct L-shape but also to detection of vehicles with indistinct L-shape. To validate the algorithm through simulation, the commonly used KITTI datasets are employed in this study. The simulation results suggest that the established model can truly reflect the real dimension and pose of a vehicle.