FedTrack: A Collaborative Target Tracking Framework Based on Adaptive Federated Learning

Tracking mobile targets is a basic task for many applications, e.g., wildlife protection, area surveillance, battlefield reconnaissance, disaster rescue, etc. In these cases, multiple edge devices are usually deployed to collect images or videos from the target area. Two mainstream tracking methodologies have been proposed to recognize and track targets from such data sources cooperatively. The centralized strategies collect data from the edge devices and thereafter run mining and learning algorithms upon such data. On the contrary, the distributed strategies implement such algorithms in a distributed manner. However, these methods incur either high transmission costs or slow convergence speeds. To this end, this paper presents a novel cooperative tracking framework (i.e., FedTrack) based on adaptive federated learning. A dual reputation mechanism has been formulated, and subsequently, an adaptive node selection algorithm has been suggested to ascertain the nodes suitable for involvement in the training process. Furthermore, a strategy for selecting the aggregation node based on capability has been developed to enhance the efficiency of aggregation. As far as we know, FedTrack is the first federated learning framework for collaborative tracking. Experimental results demonstrate that FedTrack achieves comparable or even better accuracy than state-of-the-art methods, yet needs much fewer data transmission costs and much less time consumption.