YOLO-G Abandoned Object Detection Method Combined with Gaussian Mixture Model and GhostNet

As one of the keys supporting technologies in video surveillance, recent years have witnessed a wide range of applications in leftover detection. The existing methods are not able to solve the problem of light and shadow changing. Besides, the number of parameters of the neural network in deep learning-based legacy detection rises up and the precision is not enough. This paper proposes the abandoned object detection method based on YOLO-G. The method combines background modeling of Gaussian mixture model for foreground detection. With the separation distance and time between the moving area and the static area, a suspicious static area is obtained. The frame image judged as the separation moment is sent to the deep neural network for further detection and recognition. This network model adopts the ghost module(GhostModule) in the ghost network(GhostNet) and applies it to the CSPDarknet53 backbone network. Finally, the squeeze-and-excitation network layer(SElayer) is introduced to further improve the feature extraction ability. The experimental results show that the detection accuracy of YOLO-G is 34.22%, 23.86%, 16.64%,13.19%, 8.16% and 1.41% higher than that of FCOS, SSD, RefineNet, YOLOv3, LRF and YOLOv4, and the network parameter is 22.78% lower than YOLOv4.