Extended Object Tracking With Transparency Mitigation

Zusammenfassung: We propose an efficient approach to enhance the measurement likelihood of differentiable star-convex extended target models to mitigate unlikely association processes due to self-occlusion. Our approach allows us to retain an efficient computation of likelihoods and hence the correction step for extended targets while simultaneously penalizing unlikely measurement to target associations. In combination with a filter that is able to process multimodal distributions we find that this leads to a significant increase in tracking performance, in particular during the initialization of the filter. We demonstrate an implementation based on the recently proposed Gaussian process extended target model and show how our redefined likelihood improves real-world tracking performance using automotive LIDAR data.