YOLO-Ro-KCF: a lightweight gradient-guided real-time multi-object tracking framework for embedded UAV vision systems

Robust visual multi-object tracking (MOT) from unmanned aerial vehicles (UAVs) remains a formidable challenge due to the prevalence of small-scale targets, cluttered backgrounds, motion blur, and frequent occlusions. To address these challenges in resource-constrained embedded scenarios, we propose YOLO-Ro-KCF, a novel, lightweight, and real-time MOT framework that synergistically integrates gradient-domain priors into both detection and tracking. Our core innovation lies in establishing a unified, gradient-enhanced feature representation and a closed-loop co-optimization mechanism. The YOLO-Ro detector introduces a gamma-corrected gradient magnitude map as an auxiliary input channel, substantially enhancing its discriminative power for small objects with weak textures. In parallel, the KCF-Ro tracker is augmented by fusing gradient orientation cues with Histogram of Oriented Gradients (HOG) features and incorporating a multi-scale search strategy, thereby achieving superior robustness against scale variations and partial occlusions. A dynamic fusion module adaptively reconciles detection and tracking hypotheses by leveraging spatial overlap and velocity consistency. Extensive experiments on three challenging benchmarks–VisDrone2019, UAVDT, and Anti-UAV–demonstrate that YOLO-Ro-KCF achieves state-of-the-art performance, attaining 70.6% MOTA and 72.1% HOTA on VisDrone2019 while operating at 43 FPS on an NVIDIA Jetson AGX Xavier platform. Comprehensive ablation studies, attribute-based evaluations, and cross-dataset generalization tests (yielding consistent gains of 2.1–4.5%) conclusively validate the efficacy and robustness of our approach. This work establishes a practical and deployable paradigm for high-accuracy, low-latency UAV-based visual tracking.