Real-time long-term target tracking on an ARM platform with NPU acceleration and integration into UAV line-of-sight stabilization

Abstract

This paper presents a real-time long-term target tracking algorithm optimized for ARM embedded platforms with integrated NPU acceleration. The system combines a pruned and quantized YOLOv10s detector with a NEON-optimized fDSST tracker. The two blocks are linked via an adaptive confidence index based on multi-feature fusion and a hysteresis mechanism to activate the detector only when necessary. Theoretical analysis demonstrates the boundedness of the correlation filter, the stability of the adaptive weight update process, and the exponential bounding of the probability of false state transitions. Experimental results on the Orange Pi 5 Max platform show that the system achieves an average speed of 19 FPS for detection and over 100 FPS for tracking, while maintaining stability in the presence of delay, noise, and transient occlusion. Monte-Carlo simulations and line-of-sight (LOS) stabilization simulations on a UAV rotating platform confirm a mean maximum angular error of approximately 0.006 rad and the ability to quickly re-track after target loss. The algorithm has potential applications in real-time optical surveillance, reconnaissance, and line-of-sight stabilization systems.