Onboard System for Detection, Tracking, and Georeferencing of Military Equipment in Uav Video

This paper presents an onboard hardware-software complex that identifies military vehicles in unmanned-aerial-vehicle imagery, maintains their tracks, and estimates their geographic positions. Its processing sequence covers image preparation, YOLOv7 inference, duplicate-box removal by non-maximum suppression, frame-to-frame association with SORT, ground-sampling-distance coordinate conversion, and MAVLink data exchange. Computational roles are split: the companion computer analyzes the video, whereas Pixhawk supplies telemetry and retains flight-control duties. Evaluation used four levels of scene difficulty, from controlled capture to low-light footage affected by blur, camera vibration, and occlusion. Raspberry Pi 4B processed 2.4–3.0 frames/s with a mean pipeline delay of 331–425 ms. With CUDA FP16, the GPU configuration delivered 30.2–40.8 frames/s and 24.5–33.1 ms. Across increasingly adverse conditions, mAP@0.5 changed from 0.82 to 0.65, IDF1 from 0.84 to 0.63, and coordinate MAE from 3.2 to 11.8 m. Every scenario satisfied the limits for delay, detection, tracking, and mean coordinate error. The only unmet criterion was throughput in the most demanding case: 2.4 frames/s, or 0.1 frames/s below the 2.5 frames/s target.