Object tracking with low resolution Lidar and Radar fusion, a comparison

Tracking Obstacles with low-resolution Lidar for Autonomous Driving applications is a challenging task. Learning-based models for object detection are not suitable due to the high rate of missed detections and false positives. In this work, we study two different alternatives to this approach. The first algorithm is based on Occupancy Grid detections which employs a rectangular measurement model in the tracking recursion. It is developed using the Global Nearest Neighbour (GNN) association solver and Extended Kalman Filter (EKF) estimator. A high-level fusion of the Lidar detections and Radar detections is performed. The second algorithm is developed using Extended Object Tracking (EOT) recursion, which skips the detection step entirely and utilizes fused representation of Lidar measurement points and Radar detections. It is based on Gaussian Mixture Probability Hypothesis Density (GM-PHD) filter and uses a spline measurement model. A comparative study of these algorithms with the Point Object Tracking (POT) algorithm developed with a learning-based Lidar detector is shown. The comparative study and the algorithm validation are done on the experimental data collected at the Monza Eni Circuit. (The Monza dataset will be released in conjunction with this paper).