Extended Object Tracking (EOT) Performance Comparison for Autonomous Driving Applications

Extended object tracking is an important component of autonomous driving systems, as it enables the vehicle to accurately perceive and respond to the surrounding environment. Unlike point tracking, which treats objects as single points in space, extended object tracking takes into account the shape and size of objects, as well as their motion over time. Joint Probabilistic Data Association (JPDA) and Gaussian Mixture Probability Hypothesis Density (GM-PHD) are two popular extended object tracking methods that are being used in many different engineering applications. These two algorithms have been compared and analyzed for their performance in autonomous vehicle which uses only radar data. The limited visibility of the camera under certain conditions such as foggy, sunny, or rainy weather, and its sensitivity to obstacles such as the lens being covered with rain or snow, have played an active role in not using camera sensor. Based on the results, it is shown that both methods are good at keeping track of the multiple extended objects. However, comparison of these methods shows that GM-PHD is more advantageous than JPDA in terms of Generalized Optimal Sub-Pattern Assignment (GOSPA) metric which evaluates the performance of a tracking system by measuring the difference between the estimated and true positions of the tracked object.