An Improved Object Tracking Technique for Remote Weapon Station Using Yolov۵_Deepsort_Dlib Architecture

This paper introduces an advanced tracking object architecture named DeepSORT_YOLOv۵_Dlib. Building upon the DeepSORT_YOLOv۳ framework, the study [۱] integrates the Digital Library's correlation tracker into the traditional DeepSORT_YOLOv۳ to minimize identity switches. Notably, the architecture is designed to operate in parallel, enhancing its operational speed. Experimental results indicate that the proposed approach outperforms the conventional DeepSort_YOLOv۳, showcasing reduced identity switches and increased operational speed across various video testing scenarios. The custom model employed in this study adopts a confidence threshold of ۰.۲ and an image size of ۴۱۶ x ۴۱۶, consistent with the training size. To boost detection within YOLOv۵, the model incorporates the Slicing Aided Hyper Inference (SAHI) technique. The overall inference speed in this study reaches ۳۱۴.۸fps, a notable improvement compared to Dang's ۲۱۸.۶fps. Evaluation using the COCO dataset demonstrates the model's precision at ۰.۹۸ and a recall of ۰.۸۱. Additionally, the proposed custom model exhibits a MOTA of ۰.۸۶, surpassing the benchmark's ۰.۸۳. Notably, our model achieves a significantly lower identity switch count of ۱۸۸۱ compared to the benchmark's count of ۲۲۸۸. Furthermore, it outperforms the benchmark in object detection capabilities. By incorporating SAHI inference with YOLOv۵, the study enhances detection accuracy, resulting in an overall tracking accuracy improvement from ۵۶% to ۷۹%. These findings highlight the efficacy of the proposed custom model in achieving superior performance in object tracking and detection.

Keywords:

object detection , Object Tracking , YOLO , DeepSORT , SAHI