Rao-Blackwellized Monte Carlo Data Association With Deep Metric For Object Tracking
Ajinkya Gorad, Simo Särkkä · 2023
We propose a deep Rao-Blackwellized Monte Carlo data association particle filter (DeepRBMCDA) which is a modification of the existing RBMCDA using Hungarian association. It uses YOLOv7 detected bounding box with deep ReIdentification (ReID) descriptors to track detected objects in Bayesian way. In our work, we demonstrate our performance on a diverse GMOT-40 dataset which contains sequences of varying class objects of similar appearance. We evaluate our tracker and compare its performance with state-of-the-art trackers. We obtain comparable multi object tracking accuracy (MOTA), multi object tracking precision (MOTP), localization accuracy (LocA), and multi object detection accuracy (MODA), improved mostly tracked (MT), reduced mostly lost (ML), and lowest fragmentation (Frag). We also perform the ablation study which reports highest higher order tracking accuracy (HOTA), HOTA combined LocA (HOTALocA), MOTA, identity switching (IDSW), MT, ML, Frag, and identity based F1 score (IDFI) on tracking ground-truth labels. Using particle filter for object tracking provides robustness which can be helpful in diverse dynamic tracking scenarios.