Research On Moving Target Detection Based On Deep Convolutional Neural Network Model

Moving target detection is the foundation of video analysis.The accuracy of its extraction target directly affects the accuracy of complex tasks such as behavior recognition,vehicle identification and pedestrian recognition.Therefore,it is worth to study moving target detection algorithms with high-accuracy.We mainly consider the application of deep neural network in moving target detection.By constructing and training convolutional neural network,we can extract the segmentation image of the moving target while establishing a background model.Besides,the feasibility and reliability of neural network model are analyzed and verified in the experiment.The main work is on the following parts:Firstly,to address the problem that the traditional background model can not adapt to the drastic change of the background,the convolutional neural network is introduced,and the video frame classification network is proposed.By turning the background modeling problem into an image classification problem,the network can learn the pixel value distribution of the entire background image,which uses more image information than the traditional algorithms.Thereafter,aiming at the problem that the ratio of background frames and moving target frames is not balanced,two hyperparameters are introduced in the loss function,so that the network can foucus on both the classification accuracy of the background frames and moving target frames.Secondly,since the successful application of deconvolution neural network in the field of neural network visualization,a moving target segmentation network based on deconvolution layer is proposed.The features of images are extracted layer by layer through multiple convolution layers,and then the features are restored to the original input space with the same number of deconvolution layers.Besides,the real segmentation images are added to learn the real segmentation mode of the moving target layer by layer.In addition,a square-like loss function is proposed to measure the error between the real segmentation images and the predicted segmentation images.We divide and conquer the error of the segmentation images of background frames and moving target frames.Finally,to address the problem that the features extracted by the video classification network and the moving target segmentation network in the convolutional layers are inconsistent,a multi-task learning framework is introduced,and a multi-task convolutional neural network is proposed.We can extract the segmentation image of the moving target while doing the task of establishing a background model.The network shares the parameters of two neural networks' convolutional layers,and retains other layer parameters.It can merge the two networks into a network with single-input and multi-output.In addition,the loss function is improved,and the multi-objective optimization problem is turned into a single-objective optimization problem,which improves the efficiency of the network's training.Experiments with the multi-task convolutional neural network model which is based on the video frame classification network and the moving target segmentation network and handles the background modeling problem with the moving target segmentation problem together on the real data sets show that the trained network is feasible and reliable for the detection of moving targets.Compared with other traditional mainstream detection algorithms,the false detections of the network on background pixels are fewer,and the accuracy of the extracted moving target is improved.