A Multiple Object Tracking Method Based on Optimized FairMOT

Abstract

In order to solve the issue of missed detection that is easy to occur in the multi object tracking algorithm FairMOT when the target appearance is similar to the background, and to improve the accuracy of multi-object tracking algorithm in pedestrian tracking, we proposed a pedestrian tracking algorithm termed as DA_FairMOT, based on FairMOT algorithm. At different levels of its feature extraction network DLA34, we added two self-attention modules, the spatial module and channel module. DA_FairMOT combined the two attention feature maps to further improve the representational capability of the model. In the experiment, we use the CLEAR MOT evaluation metric. As a result, the proposed DA_FairMOT algorithm improves IDP (the ID precision) by 1.59% on the MOT17 dataset, compared with the benchmark FairMOT algorithm. DA_FairMOT achieves 66.44 for MOTA, and 70.03 for IDF1.