Extended Target Tracking and Shape Estimation via Random Finite Sets

Abstract

Vehicles operating close to the ground with Light Detection and Ranging (LiDAR) pose a distinct set of challenges compared to traditional sensors such as cameras or radars. The main issues are that each target can generate hundreds of returns depending on target proximity and size, and the perceived shape of the target can vary depending on its viewing angle relative to the LiDAR. In this paper, we introduce the Occupancy Grid (OG) Gaussian Mixture (GM) Probability Hypothesis Density (PHD) filter that leverages the extended target measurements for dynamic target association and tracking. The new filter extends the GM-PHD filter to track a modified occupancy grid map representation for each target. This allows the weights of the Gaussian mixture terms to be updated in a Bayesian manner based on the similarities between the propagated target representation and the new target measurements. This filter also reconstructs an occupancy grid map representation of the tracked targets in a Bayesian manner to estimate the target shapes. The proposed filter was implemented using LiDAR data obtained from a stationary mid-tier HDL-32E Velodyne LiDAR in an urban environment. In simulations, the OG-GM-PHD filter successfully reconstructed the shape of the three tracked targets. Further, the filter tracked targets resulting in a lower Optimal Sub-Pattern Assignment error metric with up to 20% improvement and a lower cardinality estimation error compared to the traditional GM-PHD filter.