River Ship Monitoring Based on Improved Deep-Sort Algorithm

Abstract

As the economy develops rapidly, waterway transportation has gradually become an important part of the logistics industry. Aiming at the low detection and tracking accuracy of ship objects, a model for object detection and tracking of river ships was built. First, the dilated convolution was introduced into the backbone network of YOLOv3. A prediction scale of 104×104 and L2 regularization were introduced to heighten the network’s susceptibility to small objects. A target detecting model using improved YOLOv3 was constructed. Then the improved YOLOv3 was used as the detector for the deep simple online realtime tracking algorithm’s detection part. The D-IoU distance was introduced into the cascaded matching loss to build a ship tracking model based on the improved deep simple online realtime tracking algorithm. These results confirmed that the improved YOLOv3 had a detecting accuracy of 6345, a detecting time of 21.3 seconds, a recall rate of 93.25%, a missing alarm rate of 6.76%, and an average precision of 92.53%. The proposed object detection model performed the best in terms of detecting accuracy, missing and false alarm rates, and average precision indicators, with values of 87.48%, 5.14%, 12.51%, and 94.35%, respectively. The proposed ship tracking model had the highest recall rate of 64.7% and a multi-target tracking accuracy of 61.8%. This study confirms that the proposed object detection and tracking models have good performance and contribute to the intelligent development of the waterway transportation industry.