A Multiple Object Tracker with Spatiotemporal Memory Network

Target tracking is an important application of unmanned aerial vehicles (UAVs). The template is the identity of the target and has a great impact on the performance of target tracking. Most methods only keep the latest template of the target, which is intuitive and convenient but has poor ability to resist the change of target appearance, especially to reidentify a target that has disappeared for a long time. In this paper, we propose a practical multiobject tracking (MOT) method, which uses historical information of targets for better adapting to appearance variations during tracking. To preserve the spatial-temporal information of the target, we introduce a memory pool to store masked feature maps at different moments, and precise masks are generated by a segmentation network. Meanwhile, we fuse the feature maps at different moments by calculating the pixel-level similarity between the current feature map and the masked historical feature maps. Benefiting from the powerful segmentation features and the utilization of historical information, our method can generate more accurate bounding boxes of the targets. Extensive experiments and comparisons with many trackers on MOTS, MOT17, and MOT20 demonstrate that our method is competitive. The ablation study showed that the introduction of memory improves the multiobject tracking accuracy (MOTA) by 2.1.