Research on Drone Multi-Target Tracking Algorithm Based on Pseudo Depth

To address the challenges faced by drones in military and non-military operations, such as target occlusion, small-scale targets, and multiple consecutive similar targets. We adopt a detection-based tracking paradigm. First, to address the issue of low detection accuracy for small targets in the YOLOX detection algorithm, the feature pyramid network in the main network is modified, utilizing the BiFPN feature fusion method and introducing the CA attention mechanism simultaneously to enhance the network's concentration on the specific visual features of target objects and improve matching accuracy. Subsequently, in the tracking phase, through adaptive Kalman filtering and pseudo-depth segmentation strategy, optimization of matching accuracy under crowded occlusion conditions is achieved. The improved tracking algorithm demonstrated a 0.9% increase in accuracy compared to the baseline. To address the unique operational requirements in the aviation field, the algorithm proposed in this paper can meet the needs of drones, enhancing the efficiency of executing tasks such as aerial emergency rescue, disaster early warning, and search and rescue missions.