Target Recognition and Tracking Using You Only Look Once Version 5 with DeepSORT in Unmanned Aerial Vehicle Vision Systems

Nowadays, the target recognition and tracking technology in Unmanned Aerial Vehicles (UAV) vision systems has gained considerable interest due to its promising implications for various applications. However, the existing Open Flow based Tracking (OFT) system faced challenges by factors such as varying lighting conditions and occlusions. Hence, this research proposes a You Only Look Once version 5 with Deep Learning based Simple Online Real-Time Tracking (YOLOv5-DeepSORT) is proposed to handle the occlusions. The proposed YOLOv5-DeepSORT detect and track objects across frames for accurate target detection and tracking in UAV vision systems. Initially, the input data is collected from Vision-based Drone (VisDrone) dataset and then, the data augmentation is performed to expand the dataset variability. After that, YOLOv5 is employed for object detection by incorporating Transformer based Multi-Head Attention (T-MHA) as backbone to extract features. Then, GhostNet is used as feature fusion technique to minimize the computational overhead of the network. Finally, DeepSORT is introduced to track the objects by predicting object motion and associating object across frames. From the results, the proposed YOLOv5-DeepSORT achieved better results in terms of recognition accuracy (99.5%), recognition speed (33.7 vehicles/sec), and manual time (1.4s) when compared with existing UAV-Vision Based Moving Target Detection (UAV-MTD).