Probabilistic Data Association of RADAR and LiDAR Measurements for Overtaking Maneuvers in Autonomous Racing

Abstract

This thesis compares the tracking performance of estimation done using cluttered RADAR velocity measurements to estimated velocity based on LiDAR position measurements alone. The lead vehicle state estimation was done with a Probabilistic Data Association (PDA) Filter. This is a modified Kalman Filter that involves calculating the probability each detection originated from the target being tracked and using a combined measurement innovation. This prevents the track divergence that often occurs when using a filter, such as the Nearest Neighbor filter, that selects one measurement to use based on mahalanobis distance. Once the lead vehicle was tracked via the PDA filter, the state estimate was used to inform an attempt at an overtake via model predictive control. With the rise in demand for vehicles with autonomous driving capabilities, the niche of autonomous racing research has also grown. One of the main tasks to perform in a racing scenario is an overtaking maneuver. To do this successfully, the opponent’s vehicle position and velocity need to be known. This tracking is done using active perception sensors on board the ego vehicle, including LiDAR and RADAR. RADAR is capable of measuring relative velocity, where LiDAR is not, but its raw detections are noisy and cluttered. Testing of the developed autonomy pipeline was done on both simulated and real Turtle- Bot4s. The PDA filter implementation was tested with a Lincoln MKZ and Kia Optima as well. This testing was meant to validate the method of vehicle tracking for future implementation on an autonomous indy racecar. LiDAR outperformed RADAR when estimating lead vehicle position in all experiments. The velocity tracking results suggest that there may be a benefit to having a direct measurement of range rate from the RADAR sensor. This difference was not significant in the case of the TurtleBot4. The result of the full-scale vehicle tracking did indicate a greater improvement in velocity tracking at slow speeds. Highway and racing speeds were not tested.