Pedestrian LiDAR Tracking Utilizing Elliptical Model-Based MHE Through MLESAC

Pedestrian tracking using Light Detection and Ranging (LiDAR) is important to avoid pedestrians for autonomous vehicles. It is difficult, however, to measure the center position of the pedestrian directly because the point cloud data appears on the surface of the pedestrian facing the LiDAR. In addition, the arms and legs often wobble during locomotion, although the torso moves smoothly and wobbles less. Thus, in this study, to estimate the center position of the pedestrian, we approximate the torso as an ellipse model to represent the pedestrian's pose. Then, the point cloud on arms and legs that largely fluctuates is eliminated by Random Sample Consensus (RANSAC) by regarding them as outliers for the ellipse model. In addition, we propose a novel method that combines Moving Horizon Estimation (MHE) with maximum likelihood estimation sampling consensus (MLESAC) to consider the motion model of the pedestrian to prevent fitting failures. Experiments show the advantages of the proposed method.