Research On The Method Of Moving Object Detection Based On Graph Signal Processing

In the information society,information in the form of pictures and videos has shown an explosive growth.Obtaining valuable data from this vast amount of information has always been a challenging and practically applicable task.The purpose of moving object detection is to separate the valuable information in the video from the static background,i.e.,to divide the pixels in the image into two categories: foreground and background.Although deep learningbased moving object detection algorithms have achieved excellent results,recent related studies have shown that the performance of these algorithms may degrade when tested on unseen videos.At the same time,for small moving objects,existing methods are also difficult to accurately detect or cannot accurately segment the small target boundaries,thus affecting the detection effect.Therefore,our paper focuses on the following work:To address the problem that most deep learning-based moving object detection methods have significantly degraded performance in the face of unseen videos,a moving object detection method based on graph neural network with instance segmentation is proposed.To begin with,the background is extracted using the temporal median filter,and the texture,optical flow and intensity features that reflect the motion information of the instance are extracted on the result of Mask R-CNN instance segmentation.Then,the graph neural network algorithm(GraphSAGE)in graph signal processing is applied to classify the instances into background and foreground,and we improve GraphSAGE’s aggregate function and loss function to improve the accuracy of classification.Finally,our method is tested on the CDNet2014 dataset,and the results show that our method exhibits superior performance on unseen videos with excellent generalization and robustness compared to other state-of-the-art algorithms.To address the problem of motion blurring and inaccurate boundary segmentation in small object detection,a moving object detection method based on graph neural network and semantic segmentation is proposed.In this paper,a novel semantic segmentation method based on graph neural network with superpixels is designed to replace the Mask R-CNN in Part I.The aim is to improve the accuracy of the algorithm on moving small object detection task.To begin with,the image is segmented into superpixels,modeling the superpixels as nodes of the graph,and the position and color features of the superpixels are embedded as graph signals on the nodes.Next,Topological Adaptive Graph Convolutional Network(TAGCN)is used to classify the nodes,and a new superpixel penalty loss function is proposed for improving the accuracy of node classification.Then,moving small object detection on the result of semantic segmentation.Finally,our method is tested on the UAVid dataset,and the results show that our method has higher accuracy compared to other state-ofthe-art algorithms in moving small object detection.