A Novel Object Detection and Tracking Approach Using LSTM Networks and DSOD for UAV

Unmanned Aerial Vehicles (UAVs) are being used in traffic control, disaster response, and logistics. UAVs provide very detailed monitoring and object tracking, surpassing traditional surveillance systems in high-resolution data collecting and access to inaccessible regions. We develop a new object detection and tracking method utilizing LSTM networks and Deeply Supervised Object Detector in this paper. This method takes use of UAVs' mobility and high altitude. In dynamic environments, it improves object tracking. The technique relies on object labeling for precise position identifiers, Deeply Supervised Object Detector (DSOD) training on frames with improved object recognition, and LSTM for advanced processing of visual characteristics and bounding box coordinates from DSOD. Integrating the DSOD unit with the LSTM corrects the vanishing gradient problem, improving gradient propagation to capture crucial temporal correlations across frames. Occlusion, size and light variations, and camera motion are in the UAV123 dataset, which we used for our experiments. The recommended strategy outperforms all previous static camera data-based training methods in accuracy, recall, and F-scores, obtaining 97.42%, 96.13%, and 96.21%, respectively. This shows the need of training object detectors using UAV data. DSOD and LSTM provide a scalable and efficient framework for real-time object detection and tracking, improving UAV surveillance and autonomous systems. The findings suggest that deep learning methods may improve object recognition and tracking, particularly in complicated scenarios.