People Counting And Short-term Crowd Flow Forecast Under Congested Environment

Numerous occurrences of stampedes in crowded situations have indicated that crowded people management is a daunting task.Therefore,people counting becomes a crucial component in visual surveillance mainly for capacity forecasting and security monitoring.We present an occlusion sensitive method for counting crowded people in congested situations to reduce the occurrences of catastrophic stampedes and crammed events.First,we adopt the Frame Difference Method and an improved Gaussian Mixture Model to extract foreground separately,and then combine the two results together with a logical AND operation to produce a new foreground result,which is more accurate than any one of the two.Second,we spatially and evenly divide each frame into nine cell blocks for scene modeling according to multiple experiments.Third,we can get a rough number of people based on the existing algorithms.What’s more,we propose a concept which is called occlusion coefficient to solve the problem of occlusion during computing the number of people.For each cell block,the number of Harris corner points and the number of foreground pixels are necessary factors to calculate occlusion coefficient.The number of people can be obtained through the rough number and occlusion coefficient.In the experimental evaluation,our proposed method yields an improved accuracy compared with current people counting methods.As long as we get the number of people,it becomes reality that we can make accurate forecast for crowd flow.Studies have shown that the crowd flow at a given time is related to the periods before that.And this phenomenon is cyclical each day.In this paper,the wavelet neural network is designed according to the properties of crowd flow to predict the crowd flow at a certain time in the future.The wavelet neural network consists of input layer,hidden layer and output layer.There are 4 nodes in the input layer which represents the 4 time slots before the predicting time.The hidden layer has 6 nodes.The output layer has 1 node which is the forecast of crowd flow.Typically,the initial values of the network weights and wavelet basis function are randomly chosen.We use training data to train the network 240 times so that the results are closer to the real values.