Port terminal mobile recognition based on combined YOLOv5s-DeepSort

Area of Science:

• Computer Vision
• Artificial Intelligence
• Robotics

Background:

• Port and dock environments present challenges for video analysis due to scale variations, occlusion, and complex backgrounds.
• Existing object detection and tracking models struggle with reduced positioning accuracy in these dynamic settings.

Purpose of the Study:

• To develop an enhanced YOLOv5s-DeepSORT model for improved target load recognition and trajectory tracking in port and dock video images.
• To increase the adaptability and robustness of object detection and tracking systems in complex dock environments.

Main Methods:

• Integration of multi-scale convolution into YOLOv5s for enhanced multi-scale object detection.
• Incorporation of an efficient pyramid segmentation attention (EPSA) network for improved multi-scale feature fusion.
• Application of a distributed sorting loss function to address class imbalance and background variations.

Main Results:

• The enhanced YOLOv5s model achieved a 0.4% increase in mean Average Precision (mAP) due to multi-scale convolution.
• The EPSA network integration resulted in a 1.2% increase in mAP@0.5:0.95.
• Multi-target tracking accuracy (MOTA) improved by 3.1% after implementing the distributed sorting loss.
• The model demonstrated an average accuracy of 90.9% and detection accuracy of 92.2% on self-constructed datasets.

Conclusions:

• The proposed enhanced YOLOv5s-DeepSORT model significantly improves positioning accuracy and adaptability in port and dock environments.
• The combination of multi-scale convolution, EPSA network, and distributed sorting loss offers a robust solution for target recognition and tracking.
• This research provides a valuable reference for developing advanced surveillance and operational monitoring systems in maritime logistics.