Lightweight multi-target detection and tracking method for small unmanned aerial vehicles

A lightweight method for detecting and tracking small unmanned aerial vehicle (UAV) targets in complex environments, such as urban and industrial areas, is proposed. Leveraging the CenterNet target detection algorithm as its foundation, this method integrates multi-level feature fusion and a rapid spatial pyramid pooling (SPPF) structure while employing the MobileNet lightweight backbone network to ensure precise detection of small UAV targets. An enhanced DeepSORT-based multi-target tracking technique is presented to overcome the inherent instability in monitoring UAV targets with telescopic cameras. This method utilizes an adaptive noise Kalman filter (NSA Kalman Filter) for target trajectory prediction and incorporates a camera motion compensation module and BYTE target association algorithm to achieve accurate tracking of multiple UAV targets. Furthermore, a dataset for detecting and tracking small UAV targets is constructed, and the proposed algorithm is trained, tested, and validated on the embedded Jetson NX device. Experimental results demonstrate a reduction of 56.9% in average model parameter count, a 1.18% increase in mAP, and a 66.5% reduction in average computational load. With an average model size of 14.5 MB and an average processing time per frame of 36.4 ms on the Jetson NX platform, the algorithm's efficacy in accomplishing accurate identification, real-time operation, and appropriateness for deployment on edge devices with constrained computational resources is confirmed.