Experimental Evaluation of an Integrated Pipeline for Detection, Tracking, and Georeferencing of Ground Objects in Uav Video Streams

The architecture and results of an experimental study of an integrated computer system for the detection, tracking, and georeferencing of ground objects in UAV video streams are presented. The key idea is the development of a unified end-to-end computational pipeline that combines video-stream acquisition and preprocessing, object detection based on YOLOv7, temporal tracking using the SORT algorithm, estimation of spatial coordinates using UAV navigation data, and transmission of the resulting coordinate information to the flight control loop. The proposed approach is designed to operate under limited onboard computational resources and ensures coordinated interaction between computer vision, georeferencing, and navigation modules. The test results showed that the Raspberry Pi 4B platform operating in CPU-only mode achieved a processing rate of 2.4–3.0 frames per second with an end-to-end latency of 331–425 ms; in the most challenging scenario, the p95 latency reached 566 ms. In this scenario, mAP@0.5 was 0.65 and IDF1 was 0.63, while the highest mean absolute georeferencing error (MAE) across the scenarios was 11.8 m, with p95 reaching 27.0 m. The CUDA FP16 hardware-accelerated platform achieved more than 30 frames per second and an end-to-end latency of 24.5–33.1 ms in all scenarios. The results confirm the feasibility of the integrated approach and enable quantitative assessment of the relationship between detection accuracy, tracking robustness, georeferencing error, and system timing characteristics.