Methods For Moving Object Detection Based On Deep Networks

As a basic part of video processing and a crucial stage of the intelligent video analysis,the moving object detection technology has been widely used in the fields of intelligent transportation,social public safety and video compression,etc.Therefore,the research of moving object detection is of great significance.This paper mainly explores the methods for moving object detection based on deep learning.By building a model of deep networks and training it,we can realize the detection of moving objects based on supervised learning.Besides,we verify the feasibility,reliability and algorithm robustness of the deep network model.The main contributions are as follows:(1)For traditional moving object detection methods,the detection accuracy in complex scenes is not high enough,relevant scholars start from the perspective of deep learning,and achieve moving object detection by using convolutional neural networks and segmenting video frames to blocks.In view of this method is too complicated and does not consider the overall structure of video frame information,in chapter three,a moving object detection method based on deep fully convolutional networks is proposed.The method extracts the background of the video scene by the time-domain average method,and learns the nonlinear mapping between the input frame image and the background image using the deep fully convolutional networks,thereby realizing the detection of the moving object.The method can not only adapt to complex video scenes of different sizes and achieve pixel-level intensive prediction,but also requires only one forward calculation for each image in the detection process,and the method of extracting the background is simple and effectively improves the detection speed.However,because the background extraction method is simple and the deep-layer network is difficult to train,the detection accuracy of the moving object in the scene that is not involved in training is not high,and the robustness of the algorithm needs to be improved.(2)In view of the low robustness of the moving object detection method based on deep fully convolutional networks proposed in chapter 3,the fourth chapter of this paper proposes a moving object detection method based on deep residual networks and transfer learning.First of all,this method uses SuBSENSE algorithm to modify the time-domain average method of extracting the background image of the video scenes,which makes the background image extraction under complex scene more accurate.Secondly,it solves the difficulty of training deep networks and the problem of too long training time by introducing residual learning and transfer learning.Experiments show that this method not only outperforms other algorithms in the detection of moving objects,but also has a greater degree of robustness in the network model than the method based on deep fully convolutional networks in chapter 3.(3)In order to overcome the shortcomings of the original distribution of learning data in the traditional supervised learning method,the fifth chapter of this paper introduces the idea of confrontation training and proposes a detection model based on conditional generative nets.This method adds extra Ground truth information to guide the moving object detection binary image generation,and through the training,the generative model can successfully capture the true distribution of the original data.Experiments show that this method not only outperforms other mainstream moving object detection algorithms,but also,by adding a confrontation training model,the robustness of the algorithm is further improved compared to the supervised learning method proposed in the previous section.