Research On Moving Object Detection Algorithm In Complex Scene Based On Neural Network

Moving object detection performance in the complex scene is easily limited by many factors such as background motion,illumination change,shadow,etc.Currently,improves the robustness and detection performance of moving object detection algorithms in complex scene is the focus of this research field.In recent years,Convolutional Neural Network(CNN)has been successfully applied in many visual tasks,its superior performance has promoted the development of various vision fields.The application of CNN in computer vision field is becoming a research trend.This article introduces conventional moving object detection methods,further analyzes the principle of them.This paper explores the combining of CNNs and conventional methods from 3 views and proposes 3 moving object detection methods.The effectiveness of the proposed methods is verified by the experiments in the public dataset.The main work of this article is summarized as follows:However,the integrity of the foreground area mainly depends on the preprocessing,and the post-processing is mainly used to refine the detection results.There are some limitations of adopting neural networks to filter in post-processing.Therefore,from the perspective of pre-processing,this paper proposes a Moving Object Detection method based on Fully Convolutional Neural Network(MOD-FCN)using the FCN learns the foreground objects prior information.This method adopts Res Net-18 as the encoder,the network input includes the current frame and the corresponding artificial label,and then the pretrained model roughly segments the moving object area.Finally,the fine-matching moving object area is obtained by calculating the Euclidean distance between adjacent rough-matching images.The experimental results show that the MOD-FCN has a mean F-measure of 76.12% on I2 R and CDnet2014,which can adapt to a variety of complex scenes.Finally,considering the two-stage moving object detection methods have a large complexity,to improve the detection performance with small moving objects in complex scenes,this paper proposes a moving object detection method based on Deep Frame Difference Convolutional Neural Network(DFDNet).DFDNet is a one-stage moving object detection method.The network is composed of Difference Net and Appearance Net,and it can segment fine-matching foreground object regions without post-processing.Difference Net is used to learn the differences between two consecutive frames and capture temporal information from the inputs(t frame and t +1 frame).Appearance Net is used to extract spatial information from the input(t frame)and fuse it with temporal information.Besides,the multi-scale spatial information is retained by multi-scale feature map fusion and stepwise up-sampling to improve the sensitivity to small objects.The experimental results show that the mean F-measure of the DFDNet reaches 82.54% on the I2 R and CDnet2014,which can effectively segment the foreground region,not only adapt to various complex scenes but also significantly improve the detection performance of small moving objects.This article respectively explores post-processing,pre-processing and deeply network model three research perspective,respectively proposes a moving object detection method based Gaussian Mixture Model and BP neural network,a moving object detection method based on the Fully Convolutional Neural Network and a Deep Frame Difference Convolutional Neural Network for moving object detection.The performance of methods are proved by experiments,which provides a new research idea for moving object detection algorithm in complex scenes.