Intelligent video surveillance of sand mining based on object detection and tracking

Abstract

In order to improve the regulatory capability of sand mining and make full use of the video resources collected by video surveillance systems, an object detection and tracking algorithm that combines YOLOv4 and Kalman filter is proposed to identify sand dredgers. Initially, an object detection dataset of common ships is finely established through camera capture, manual and automatic annotation, and data augmentation. Then, the obtained dataset is used to train YOLOv4. The results show that the trained YOLOv4 achieves a mean Average Precision of 92.26% over all classes. Although Kalman filter does not increase the precision of YOLOv4, it supplements the association information between images in an frame stream, so as to solve the target loss problem of YOLOv4 caused by occlusion. Finally, the proposed algorithm is deployed in edge computing terminal devices to form an intelligent video surveillance system for sand dredgers. The intelligent system realizes the 24-hour real-time monitoring of sand mining, which can help to stop illegal sand mining in time, and provide strong technical support for relevant departments to fulfill their management obligations.