Study On Moving Shadow Detection And Object Recognition

With the development of science and technology and the improvement of social economy,the mobility of the population has been increasing.A large amount of floating population brings challenges to public security.Traditional video surveillance system relies on human to monitor the scene,which cannot give full play to the role of video surveillance.In recent years,computer vision,machine learning and some corresponding technology have developed rapidly,which makes intelligent video surveillance system widely used in many public places.Compared with traditional methods,intelligent methods not only reduce the workload of people and save cost,but also improve the ability of data processing and anomaly detection.Combined with current advanced technology,intelligent video surveillance system will play an important role in building smart city,safe city and so on.Due to the diversity and complexity of surveillance scene,intelligent video surveillance still faces many problems and challenges.This paper mainly focuses on moving shadow detection and moving target recognition.Its main work is as follows:(1)The formation principles and properties of moving shadow are analyzed.Various types of moving shadow detection algorithms are introduced.Some classical moving target detection methods are introduced,the algorithm based on codebook is described in detail.The frequently-used features and classifiers in target classification are introduced.(2)A moving shadow detection algorithm based on Haar Local Binary Pattern(HLBP)feature is presented.The algorithm extracts HLBP feature of detected moving regions and corresponding background regions separately.It does not need threshold selection,image block histogram statistics.And it uses Manhattan Distance to measure the difference of HLBP feature between the two regions to get a map about texture difference,then segments texture difference map to get a mask of moving objects.Combined with color space and texture information,a moving shadow algorithm based on random forest classifier is presented.The method reduces the hypothesis of the illumination and reflection properties of the scene,and avoids the setting of parameters.Random forest classifier is trained to predict whether the pixel is shadow or not.Experimental results show that the method can achieve good results and have generalization ability in various indoor and outdoor scenes.(3)A moving target classification algorithm based on HOG-HLBP feature is proposed.The algorithm combines Histogram of Oriented Gradient(HOG)and HLBP to use their comprehensive description ability of target contour and texture.Support Vector Machines(SVM)is employed to predict label.Experimental results show the effectiveness of the method.A multi-target tracking method based on detection is implemented.The appearance information is considered in cost function.Experimental results show the method can deal with some occlusion scenes.Combined with motion detection,moving shadow detection and multi-target tracking,surveillance video moving target classification is realized.