UAV-STNet: A Hybrid U-Net–Transformer–LSTM Architecture for Real-Time Object Detection in UAV Video Surveillance

Abstract

It is difficult to perform real time object detection in Uncrewed Aerial Vehicle(UAV) based video surveillance because of the dynamic movement of camera, scale, occlusion, variation of illumination and limited power availability on-board of the computer. Purpose: In this paper of researches, the author suggests the proposed UAV Spatial-Temporal Network(UAV-STNet), which is a hybrid model of spatio-temporal deep learning model that is expected to improve the accuracy of the detection, balance of time, and real-time performance. Approaches: techniques: The proposed framework incorporates U-Net representing the system of total encoder involving extraction of multiscale space features mode, Transformer attention module which will involve global contextual modeling and Long Short-Term Memory(LSTM) network which will involve learning the short-term dependency in sequential frames. The model has been lives with and examined against a personalised UAV video dataset using a trade mark assortment of generally utilised measures of detection, including precision, recollection, mAPat 0.5 and inference velocity (FPS). Findings: UAV-STNet gave 94.5% precision, 93.1% recall and 93.2% mAP @0.5 and also 45 FPS. Better accuracy-tolerance efficient state is observed in comparative comparison with SSD, YOLOv3 and faster R-CNN which are evidently small object in motion affected scenes. Conclusions: The spatial, contextual and temporal modelling applied as a single-stop end-to-end architecture shows a powerful and computationally efficient solution of the topic of detecting real-time UAV video items, that brings benefits in the stability and reliability of intelligent surveillance systems using aerial vehicles.