Deep learning method for UAV detection and flight path forecasting in real-time

Abstract

The paper is devoted to the creation and approbation of the neural network method for detecting and forecasting the trajectories of small UAVs in a video stream that integrates detection, tracking, and transformer prediction into a single high-performance pipeline. The architecture combines the YOLOv12 model for object localization, the BYTETrack algorithm for identity maintenance, and the modified position-coded Transformer for short-term motion forecasting based on coordinate history. The system generates continuous trajectories in conditions of a changing background, maneuvering, and partial overlaps and provides an average RMSE prediction error of 12.66 pixels, which is interpreted as approximately 0.25–0.32 m in the physical space of the scene. The obtained values of MOTA = 0.79 and IDF1 = 0.64 confirm the stability of tracking and preservation of object identity. On the Anti-UAV benchmark of 593,802 frames of real surveillance video, the YOLOv12 detector within the pipeline achieves mAP50 = 91.04% and mAP50–95 = 62.76% on the test split, running at 55 FPS on GPU and 13 FPS on CPU, while the BYTETrack-based tracker reaches 332 FPS on CPU. The experiments are carried out on the visible-spectrum subset of the Anti-UAV dataset and cover both offline processing of recorded video and online processing of live camera streams. Comparative analysis shows that the proposed approach significantly outperforms statistical models (IMM/Kalman), reducing the forecast error by an order of magnitude. The identified limitations relate to work in two-dimensional space with a fixed observation angle and a short forecasting horizon; extending the model to mobile platforms, multidimensional coordinates, and behavioral forecasting are identified as promising areas for further work. The results of the study confirm the effectiveness of combining transform forecasting with modern detection and tracking methods for building situational awareness systems, airspace monitoring, and countering unauthorized use of UAVs.