Machine Learning Based Study On Automatic Pedestrian Counter In Video Sequence

Intelligent video supervising technology has increasingly become the important technical means of public security management,and automatic passenger counting is one of the important contents of intelligent video supervising,it is of great significance for social peace.Because of the large information in the video sequence,method existing can not get accuracy and real-time performance at the same time.The main research content and results are as follows:An improved multi-target tracking method is presented for solving the problem that tracking effect turns bad when objects have a higher moving speed.This method does Kalman predictions on detected objects after extracting the moving objects by the method of background subtraction,and the results would be taken to iterate as the initial position in Mean-shift search.Then,the result would be regarded as the observed value in Kalman filter correction phase,by which we can finish the update of the Kalman filter.Apart from it,we introduced the occlusion factor to have a judge on objects' occlusion condition,by which we can achieved auto-adaption processing on occlusion.Experiment results indicate that this method can reduce cases of missed tracking effectively,has a good robustness for the objects with fast speed,and can achieve correlation matching for the same object in any different frames.Aiming to the problem of complexity operation which was caused by the high dimension of human features,we proposed a feature reduction method based on rough set.Extracting a series of representative features,then reducing the features through the method of attribute reduction in rough set.Experiment results indicate that this method can reduce the cost of recognition,and meet real-time requirements.In addition,for the extracted pedestrian features,we studied the methods based on machine learning to count the numbers of people,and used a optimized BP(back propagation)algorithm which has a self-adaptive momentum factor.This method introduced a self-adaptive momentum factor to update weight values in each layer and finish the back propagation.Experiment results indicate that we can refine the instability in BP algorithm which was caused by the wrong picked value on momentum factor,and had a low time complexity in this method.The research technique of automatic passenger counting in the video sequence in this paper can make full use of the information in the surveillance video,achieve the automatic detection,tracking,and get the accurate number of pedestrian to eliminate the potential security risks of public place which has high density population.It is of great importance to the management of public and the prevention of crowd catastrophe.