Research On Method Of Pedestrian Detection And Tracking In Video

In daily life,the technology of pedestrian detecting and tracking in video has a very wide range of applicationsm,such as intelligent monitoring,auxiliary driving.Only the pedestrian is detected and tracked in real-time and accurately,the nachine can understand the pedestrian's action and make a corresponding instruction.The idea of this thesis is to detect the moving objects first,then select the pedestrian from the moving objects,and finally the pedestrian is tracked.The contents of this paper are as follows:(1)Moving objects detecting.The background modeling method about moving objects detection is analyzed in this paper.Experiments are researched to the mixed Gauss background modeling,the Codebook modeling and the ViBe(Visual Background Extractor)background modeling.The results indicate that the mixed Gauss background modeling has a poor performance,the Codebook modeling is ordinary,and the ViBe is good.On this account,the ViBe is applied and we make 3 points of improvements which are background template initialization,sampling range and the threshold of pixel classification.The results show that the improved ViBe has a better performance than the original ViBe algorithm,the outline of the objects are more complete and fewer background noise.(2)Pedestrian detection.The moving pedestrian is selected from the moving objects for tracking.The pedestrian discrimination model is trained with the SVM(Support Vector Machine)algorithm to classify the moving objects.And the HOG(Histogram of Oriented Gradient)feature is used as the classification characteristics.The results show that the pedestrian recognition model's accuracy rate is 0.947 and can meet the requirement of pedestrian detection.(3)Pedestrian tracking.The pedestrian is tracked after the detecting.The Mean-shift algorithm,the ASMS(Adaptive Scale Mean-shift)algorithm and the KCF(Kernelized Correlation Filter)algorithm are researched deeply.The algorithm's accuracy and processing time are compared through experiments which are target occlusion experiment,motion blur experiment and scale change experiment.The experiment's results show that Mean-shift's accuracy is poor and it's minimum processing time is 65.1ms.The ASMS's accuracy is moderate and it's minimum processing time is 33.4ms.The KCF's accuracy and processing time both are the best among the three algorithms,and it's minimum processing time is 30.2ms.The KCF algorithm is selected to track the pedestrian.Through pedestrian detecting and tracking experiments,it is shown that the research scheme is correct and can track the single target pedestrian in real time and accurately.