Scholarlink

<p>The rise of urban traffic has resulted in overcrowding of vehicles and road casualties. Sustainable traffic advancement based on vehicle detection and tracking efficient road mobility solutions for congested roadways. The proposed research incorporates modifications to version 5 of You Only Look Once (YOLOv5) for detection along with the refined Deep Simple Online and Real-time Tracking (Deep SORT) for vehicle tracking. A new multi-type vehicle detection dataset called Vehicle Identification aNd Classification (VINC) is introduced. The structure of the proposed YOLOv5s detection model is enhanced with attention mechanism to small object detection. The length-to-width ratio utilized by the Kalman filtering algorithm is replaced only by width in the vehicle prediction box improving Deep SORT’s state prediction rate. The proposed model achieves a precision, recall, and mAP rate of 97.2%, 92.4%, and 77.2% respectively on the VINC dataset exhibiting equivalent detection accuracy in detecting both small and large vehicles. Our approach achieves a multi-object tracking precision (MOTP) of 88.54% and a multi-object tracking accuracy (MOTA) of 79.32% on the VisDrone dataset. These results demonstrate the proposed system’s ability for precise real-time vehicle monitoring by balancing accuracy and inference time.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Sustainable Real-Time Urban Mobility with Attention-Based YOLOv5s and Refined Deep SORT Algorithm

  • 期刊论文

  • 发表日期：

    2025年10月6日

查看全文
  • B. Cynthia Sherin,
  • E. Poovammal

摘要

The rise of urban traffic has resulted in overcrowding of vehicles and road casualties. Sustainable traffic advancement based on vehicle detection and tracking efficient road mobility solutions for congested roadways. The proposed research incorporates modifications to version 5 of You Only Look Once (YOLOv5) for detection along with the refined Deep Simple Online and Real-time Tracking (Deep SORT) for vehicle tracking. A new multi-type vehicle detection dataset called Vehicle Identification aNd Classification (VINC) is introduced. The structure of the proposed YOLOv5s detection model is enhanced with attention mechanism to small object detection. The length-to-width ratio utilized by the Kalman filtering algorithm is replaced only by width in the vehicle prediction box improving Deep SORT’s state prediction rate. The proposed model achieves a precision, recall, and mAP rate of 97.2%, 92.4%, and 77.2% respectively on the VINC dataset exhibiting equivalent detection accuracy in detecting both small and large vehicles. Our approach achieves a multi-object tracking precision (MOTP) of 88.54% and a multi-object tracking accuracy (MOTA) of 79.32% on the VisDrone dataset. These results demonstrate the proposed system’s ability for precise real-time vehicle monitoring by balancing accuracy and inference time.