Tracking 3D LIDAR Point Clouds Using Extended Kalman Filters in KITTI Driving Sequences

Abstract

In this paper, we present a novel approach to track recognized 3D vehicle point clouds from LIDAR scans. This technique is based on tracing the 3D anchor boxes of these vehicles, recognized though convolutional neural networks (CNNs). Exploiting the 3D CNNs detection of vehicles and persons, and the Extended Kalman Filters (EKF) two-steps process for prediction and update, the proposed scheme guarantees the awareness of moving detected objects and improves the perception of Autonomous Vehicles (AV). The proposed scheme reduces the usage of the expensive detection process of feeding the point cloud to CNN by tracking the 3D rectangular coordinates containing already detected objects from early Velodyne scans of the driving sequences. The testing of the proposed method on the well-known KITTI dataset, featuring LIDAR scans of realistic vehicular environments. Results show the merits of the proposed scheme in achieving high tracking accuracy.