Model training parameter configuration.

The Procapra przewalskii, plays a vital role in sustaining the ecological balance within its habitat, yet it faces significant threats from environmental degradation and illegal poaching activities. In response to this urgent conservation need, this article proposes a multi-object tracking (MOT) method for unmanned aerial vehicle (UAV). Initially, the approach utilizes a modified YOLOv7 network, which incorporates Group-Selective Convolution (GSConv) in its Neck component, effectively enhancing the network’s ability to preserve detailed information while simultaneously reducing the computational load. Subsequently, the Content-Aware ReAssembly of Features (CARAFE), an innovative feature upscaling method, replaces the conventional nearest neighbor interpolation to minimize the loss of critical feature data during image processing. In the tracking phase, the Deep SORT algorithm is expanded with a proprietary UAV camera motion compensation (CMC) module that eliminates the impact of UAV camera jitters. Moreover, the system has incorporated a confidence optimization strategy (COS) that enhances the tracking performance especially when the individuals are partially or fully obscured. The method has been tested on Procapra przewalskii and shown to be effective. The results show the gains in tracking metrics where the method achieved improvements of 7.0% in MOTA, 3.7% in MOTP, and 5.8% in IDF1 score compared to the traditional Deep SORT model. Improved tracking methods can alleviate the impact of occlusion and rapid movement of UAV on tracking, thereby more accurately monitoring the status of each Procapra przewalskii and protecting it. Also, the efficiency in the multi-target tracking achieved through the use of this system is sufficient for the operational demands of UAV-based wildlife monitoring, thus being a reliable tool in wildlife conservation where accurate and efficient wildlife tracking is desired.