Research On Moving Object Detection Method Based On Instance Segmentation

In recent years,with the popularization and development of surveillance video technology,moving object detection technology has been widely used in industrial production,security monitoring and traffic management,and other fields.However,the variability and complexity of actual monitoring scenes bring many difficulties and challenges to moving object detection algorithms,such as dynamic background,camera jitter,shadows,etc.,which will affect the accuracy and robustness of the algorithm.Therefore,research on designing a moving object detection algorithm with good real-time performance and strong robustness has become the focus of current technical research.Instance segmentation methods based on deep learning exhibit strong robustness and can accurately segment the contours of each object,so they are an effective solution.This paper focuses on the instance segmentation method based on deep learning and conducts in-depth research on the moving object detection technology in surveillance video.The main research content is as follows:(1)On the mask branch of conditional convolution instance segmentation,the semantic difference between the multi-scale features to be fused is too significant,which is not conducive to the problem of network parameter learning.A conditional convolutional instance segmentation model based on multi-scale progressive fusion is proposed.First of all,in the mask branch,to better fuse multi-scale features,this paper adopts a multi-scale progressive fusion strategy to learn the feature information of each layer to obtain richer mask features.Secondly,to solve the problem that the object center is sometimes not on the actual object,this paper introduces the massness,which uses the object’s center of mass to represent the object center position to locate the object position more accurately.The experimental results show that on the COCO dataset,the segmentation accuracy of the instance segmentation model MPCInst proposed in this paper is 39.3 m AP,which is better than the Cond Inst model,indicating that the improved Cond Inst is effective.(2)Aiming at the problem of a large number of false positive pixels and false negative pixels in the detection results of the traditional moving object detection algorithm in the actual surveillance video,a moving object detection algorithm based on instance segmentation is proposed.First,an instance segmentation model and a moving object detection algorithm generate binary instance masks and a coarse motion mask,respectively.Then,high-quality and low-quality binary instance masks are obtained by thresholding the binary instance masks.Next,the high-quality and low-quality binary instance masks are fused with the coarse motion mask by instance-wise fusion and instance-union fusion,respectively.Finally,the output results of instance-wise fusion and instance-union fusion are combined to obtain the final accurate motion foreground mask.The experimental results show that on the CDNet-2014 dataset,compared with the benchmark moving object detection algorithm Vi Be,the comprehensive F-Measure of the Inst MD algorithm combined with the instance segmentation model MPCInst and Vi Be is 0.7991,an increase of 29%.When combined with more accurate benchmark algorithms(moving object detection algorithm and instance segmentation model),compared with the benchmark moving object detection algorithm Su BSENSE,the comprehensive F-Measure of the Inst MD algorithm combined with the instance segmentation model CBNetv2 and Su BSENSE is 0.8244,an improvement of 13%.When combining real-time benchmark algorithms,the Inst MD algorithm was implemented using the instance segmentation model YOLACT++ and the Vi Be algorithm to achieve real-time detection at 30 FPS on a video sequence with a resolution of 320 × 240.Therefore,the Inst MD algorithm proposed in this paper can significantly improve the accuracy of the benchmark moving object detection algorithm and has robust scalability.It can improve its performance with a more accurate and faster benchmark algorithm.