Research On Background Modeling Algorithm Based On Fusion Color Data And Depth Data

After entering the twenty-first century,social public security incidents occur frequently in the world.Causing a serious impact on the people’s life and psychology,promoting the public to strengthen the awareness of safety concerns.As a result,many countries begin to study intelligent video surveillance system to liberate the manual labor,which can provide accurate,real-time warning information.Along with the rapid spread of robot technology,it is worthy of further study for robots to identify the target correctly and avoid obstacles in the road.Moving target detection is an important module in the application of machine vision system,which provides critical services in both of the above systems.While background modeling algorithm is a technology to achieve the detection of moving targets.This paper first analyzes the common problems of background modeling algorithm in the color space: changes in external lighting cause scene brightness changes;the color of the foreground is similar to the background area;the shadow of the target causes the change of background brightness.These optical problems can be resolved by other sensor data,this paper uses the depth data.While the use of deep data introduces a new problem at the same time: the ADO area.In order to solve the above problems,we have built the background model by combining the depth data and color data in this paper.Firstly,we improved the original mixture of Gaussians model,and built double background model,one is color-based the other is depth-based.The model is optimized according to the ADO area in the depth map to reduce the interference of numerical mutation to the model,which causes model to fail to converge.We update the background model based on the fusion results by using two update rate values: strong or weak,which enhances the stability of the model,while improving the ability to adapt the change of background.Then use the new decision algorithm to fuse the output of the dual background model.Finally,compared with the experimental results of other algorithms,it is found that the detection rate of this algorithm is improved.In order to improve the accuracy of the fusion results of the dual background model and avoid the simple binary operation,this paper proposes a fusion algorithm framework based on the classifier.Extracts the color-based and depth-based features from the RGBD datasets.Candidates features used in this paper are HSV values,edge features,LBP features,and so on.Using this data to complete the training of the classifier,the classifier is random forest in our method.The output of the dual background model is divided into different regions,and the new results are recalculated by classifier.In order to improve the accuracy of the fusion results of the dual background model and avoid the simple binary operation,this paper proposes a fusion algorithm framework based on the classifier.Extracts color-based and depth-based basic features from the annotated RGB-D datasets.Candidates used in this paper are HSV values,edge features,LBP features,and so on.Using this data to complete the training of the classifier,the classifier used in this paper is a random forest.The output of the dual background model is divided into different regions,and the new results are recalculated using the classifier.This framework can be applied to a variety of background modeling algorithms,MOG algorithm,AdaptiveMedian algorithm,FuzzySOM algorithm,StaticFramDifference algorithm.Finally,the fusion results are quantitatively analyzed,and the results of other experiments are compared.It is found that the framework can improve the detection rate of the foreground target.