Research On The Frontier Technology Of Foreground Detection

In today's society,with the development of sensors and computer vision technology,people have more needs and devote more research works on video surveillance automation and intelligent technology.As a basic technology in the field of computer vision,foreground detection aims to separate moving foreground objects from relatively static background scenes.It is the first problem in computer vision analysis and understanding tasks.Currently,due to the complexity of application scenarios,the performance of traditional foreground detection algorithms faces a bottleneck.Improving algorithm accuracy,enhancing algorithm real-time application capabilities,and improving algorithm robustness are the research focuses of foreground detection technology.This paper first studies the foreground detection algorithm based on the fusion of color and depth information for application scenarios such as shadow and color camouflage.Secondly,this paper proposes a foreground detection framework based on matrix decomposition and mathematical optimization to solve the application scenarios of dynamic background such as lighting,water ripples and swaying leaves,and proposes solutions from three aspects for the problem of computing resource consumption under this framework.Finally,this paper studies the detection framework based on deep learning neural network,improves it from three aspects of optimization training data,lightweight network structure,and multi-scale feature map module parameter optimization,and verifies the detection efficiency through experiments.The main work and research innovations of this article are as follows,1?In order to improve the accuracy and robustness of the traditional foreground detection algorithm in shadow and color camouflage scenes,this paper improves the performance of the foreground detection algorithm through the fusion of color information and depth information.The background model uses the traditional algorithm with good real-time performance,high detection rate,and strong adaptability-Sub SENSE(Self-Balanced SENsitivity SEgmenter).The detection process calculates the four state matrices of background confidence matrix,noise accumulation matrix,distance threshold matrix,and update probability matrix to perform local dynamic threshold setting and automatic pixel classification.However,the algorithm has a disadvantage that when the foreground and background are similar,the update strategy of the algorithm would reduce the foreground detection rate.This paper proposes a fusion method of depth information and color information to handle shadow,color camouflage,and depth camouflage situations,which can improve the foreground detection rate.Experiments prove that the fusion strategy can greatly improve the detection rate.2?According to the characteristics of the video foreground target,the optimization function of the sparse matrix decomposition optimization algorithm(subspace learning)is improved,and the variable in the target function is updated and solved using the fast alternating optimization algorithm which greatly reduces the calculation time complexity.Secondly,in view of the consumption of computing resources and storage resources,a frame screening algorithm is proposed.By screening redundant frames for long videos,the dimension of data is greatly reduced,which can effectively reduce the amount of data involved in matrix decomposition.Finally,considering the batch processing has high requirements on storage resources and computing resources,a dictionary-based online optimization algorithm is proposed to substitute the traditional batch processing of matrix decomposition.Compared to batch operations,the online strategy really solves the problem of the algorithm's huge consumption of storage space and resources.3?Convolutional neural networks have powerful feature expression and data fitting capabilities in many computer vision tasks.This paper studies the foreground detection framework based on deep convolutional neural network.Firstly,for the problem of imbalanced training data categories,this paper proposes a filter and balance strategy to reduce the number of invalid training samples.The experiment in this paper verifies that the method can use a small number of manually labeled samples to model the data set without lowering detection efficiency.Secondly,lightweight model is proposed for reducing model training time and large parameter scale.Finally,combining the training sample screening equalization strategy and the lightweight network,a suitable multi-scale feature extraction module is proposed.This algorithm can improve the detection effect while greatly reducing the size of the parameter model and improving the detection speed.Finally,the algorithm verifies its effectiveness on the data set.