Density Enhancement-Based Long-Range Pedestrian Detection Using 3-D Range Data

The ability to perform long-range pedestrian detection is essential for autonomous vehicles. However, for 3-D LIDAR, an object's point cloud becomes sparse when it is away, directly affecting its detection as a result. In this paper, a novel density enhancement method is proposed to improve the quality of a sparse point cloud. The input of the method is an object's raw point cloud; first, a high-quality local coordinate system of the point cloud is built using a new evaluation metric, and then radial basis function (RBF)-based interpolation is performed based on the local coordinate system. Finally, a resampling algorithm is used to generate a new point cloud that not only meets a density requirement but also fits the object's geometric shape. Novel features of our method are its evaluation metric of a local coordinate system and method to choose a good shape parameter and kernel in RBF-based interpolation step. The effectiveness of this method is demonstrated using naturalistic data and three experiments.