Research On Multi-scale Multi-modal Moving Object Detection Method Based On Low-rank Sparse Factorization

In computer vision,moving object detection is a basic problem.is to Extracting moving objects from continuous video frames is its main task.It has a very valuable application in vehicle navigation,scene understanding,industrial robot technology,intelligent transportation and other fields.In the past few decades,experts and scholars have made remarkable achievements in the detection of moving object.However,in practical applications,due to the uncertainty of the environment,it will be sensitive to the changes of the scene and light,and will also be affected by bad weather factors such as occlusion,rain and snow,moving object detection is still a challenging task.Recently,the low-rank and sparse factorization framework has been well applied in moving object detection,and with the rapid development of thermal infrared,depth sensor and other technologies,multimodal moving object detection has also attracted much attention.Therefore,this paper studies a multi-scale and multimodal moving object detection method based on the low-rank and sparse factorization.Researched and explored the following issues,as follows:(1)Aiming at the problems of rough boundary,incompleteness and sensitivity to noise in the detection results of current moving object detection methods,a moving object detection method based on multi-scale low-rank and sparse factorization is proposed.In the low-rank and sparse factorization framework,the matrix representing the video frame sequence is decomposed into low rank matrix and sparse matrix.On this basis,this paper proposes to implement the appearance consistency and space compactness constraints under the multi-scale structure for the foreground,so as to better restore the background and the foreground.At the same time,based on the diversity of superpixel segmentation at different scales,this paper proposes to introduce space compactness by implementing the consistency between each pixel in the superpixel at different scales.In addition,pixels with the same appearance are easy to be divided into a super pixel block,so this strategy can promote appearance consistency at the same time.In order to improve the efficiency of the algorithm,introduce an alternative definition of the nuclear-norm with a efficient decomposition strategy.Through the analysis of experimental results on the common data set GTFD and CDnet,it can be obtained that the method presented in this paper has greater performance advantages than the comparison experimental method.(2)Aiming at the problem of poor detection results in bad weather,night and other atmospheric visibility scenarios,this paper proposes a multi-mode algorithm based on low-rank sparse factorization multi-scale structure for the detection of moving targets.The imaging advantages of visible light camera are clear spatial structure and texture,high resolution and rich color features.However,its disadvantages are also obvious.The imaging results are easily affected by light.For example,in the case of low visibility,the imaging quality is poor,and at night,it cannot be imaged.However,the image of thermal infrared camera comes from the infrared thermal radiation of the target itself,which is characterized by normal imaging in the case of poor visibility such as rain,snow,haze and night.Therefore,when the image quality of the visible camera is poor,the thermal infrared camera can observe the target normally.However,thermal infrared imaging also has its own disadvantages,such as thermal profile and low resolution.Therefore,this paper proposes a more robust multi-modal moving target detection algorithm by using the complementary information among multiple modes.Based on the low rank sparse decomposition model,an adaptive weight weight is introduced to capture the cooperation and heterogeneity among different modes in multimodal foreground detection.In addition,the appearance consistency and space compactness constraints under multi-scale structure are integrated in each mode to improve the accuracy of detection results.Through the analysis of experimental results on the common multi-mode data set GTFD,it can be obtained that the proposed method has greater performance advantages than the comparison experimental method.