| Background subtraction is one of the most important research topics in computer vision,where the objective is to solve the problems of how to effectively construct the background model of the scene,and accurately detect the moving foreground objects in the scene.Therefore,the background subtraction methods have good applica-tion prospect and high practical value in the areas of intelligent video surveillance,human-machine interaction and video coding.However,the background subtraction methods also confront many problems and challenges(such as illumination changes and dynamic backgrounds etc),and how to effectively deal with these problems and challenges has became an urgent problem.In the meanwhile,the background subtraction methods based on deep neural networks have gradually attracted much attentions in recent years.Owing to the stroing representation capability of the deep neural networks which can effectively characterize the scene background and foreground objects,these methods have achieved good performance for background subtraction,and can effectively address some problems and challenges in current background subtraction methods.Therefore,it is of great practical significance and scientific research value to address and the problems and challenges in the back-ground subtraction methods,and to explore the background subtraction methods based on the deep neural network.The main work of this article is given as follows:In view of the problems and challenges confronted by the current background sub-traction methods,this thesis proposes a background subtraction method based on multi-strategy fusion.On the basis of ViBe method,the proposed method not only develops a new strategy of foreground object detection,but also introduces the strat-egy of the foreground object aperture procesing based on bilateral aperture detec-tion,the strategy of foreground noise detection based on edges detection,the strategy of ghost region detection based on the statistic of ghost count value and temporal-spatial discontinuity.The experimental results obtained by the proposed method on the CDnet-2014 dataset show that the proposed method is higher than that of the ViBe method by 15.5%,which shows that the high robustness of the proposed method.In the meanwhile,this method obtains the average processing speed of 44 frames per second on the dataset,which verifies the high execution efficiency of the proposed method.In addition,this thesis also proposes a background subtraction method based on multi-scale cascaded neural network.The proposed method analyses the drawbacks of the CascadeCNN method in the design of network model structure,and a multi-scale cascade convolutional neural network model with a simpler model structure is proposed.In the meanwhile,this thesis proposes a network model training strategy based on the balance of positive and negative samples,which enables the network to perform efficient model training based on small sample data and obtain a more robust background subtraction performance.The experimental results on the CDnet-2014 dataset show that,the neural network models of the CascadeCNN method and the.proposed method are trained by using of the 5 frames,10 frames and 20 frames train-ing samples respectively,the experimental results obtained by the proposed method are higher than that of CascadeCNN method by 4.99%,2.83%,1.30%respectively.It shows that the proposed method achieves better performance under the condition of using small sample training data. |
