Real-time tracking algorithm for unmanned dynamic targets based on lightweight detection

A method for the tracking of targets of UAVs in low-altitude situations is proposed, aimed at solving the problems of low accuracy, inadequate real-time efficiency, and huge model size throughout actual time detection of drones, as well as the unexpected disappearance of the target during tracking. To reduce the number of parameter values while creating a lightweight network architecture, the RepViT model is utilized as the backbone network within target detection. To increase the sensing field and decrease the loss of data, the CARAFE user is used in place of the nearest-neighbor interpolation up-sampling module. Motion target trajectory prediction is used in target tracking to improve tracking accuracy by using past data to predict the future motion of the target. The verification set’s mean accuracy (mAP50) is 1.9% higher than the YOLOv5s model, and the inference speed is similarly increased, according to experiments done on the VisDrone dataset. As demonstrated by the tracking trials, the system can safely and practically handle UAV monitoring duties by accurately tracking the subject and preventing impediments while taking flight safety into consideration.