Multi-Target Tracking with Multiple Unmanned Aerial Vehicles Based on Information Fusion
Pengnian Wu,Yixuan Li,Dong Xue
DOI: https://doi.org/10.3390/drones8120704
IF: 5.532
2024-01-01
Drones
Abstract:In high-altitude scenarios, targets tend to occupy a small number of pixels within the UAV’s field of view, resulting in substantial errors when identity recognition is attempted based solely on appearance features during multi-UAV joint tracking. Existing methodologies typically propose projecting multi-view data onto a single plane and leveraging distance information for identity association; however, their accuracy remains low as they are contingent on one-dimensional target information. To address this limitation, this paper introduces the UAVST-HM (UAV Swarm Tracking in High-altitude scenarios for Multiple targets) model, specifically designed to handle the characteristics of targets in the field of view of multiple UAVs at high altitudes. Initially, we develop techniques for extracting targets’ appearance, geometric, and distribution features. Subsequently, adaptive weights, calculated based on the mean of the respective features, are devised to amalgamate these diverse features, thereby constructing a cost matrix for cross-view target identity matching. This matrix is processed through the Hungarian algorithm, and multi-view target identity association is ultimately achieved via threshold filtering. On the MDMT dataset, our method enhances the MDA indicator, which assesses cross-view target identity matching, by 1.78 percentage points compared to the current state of the art. This significant enhancement substantially improves the overall efficacy of multi-UAV joint visual tracking from a high-altitude perspective.