Crowd Density Analysis With Convolutional Neural Network

In recent decade,the increased crowd activities in commercial district,plaza,transport station and other public places for business,recreational and religion promoted an economic boom,but also brought enormous security problems.While people gather to a dense crowd,it is easy to cause jam,stampede accident,social safety event and public health event.Thus,crowd density analysis with video surveillance becomes more popular and important in public security.The early detection of density distribution and crowd behaviors are also important for crowd control and emergency plan.Crowd density analysis includes crowd counting,crowd density distribution estimation and crowd abnormal behavior detection.Crowd counting is for global density estimation,we propose a multi-scale fusion and recursive convolutional neural network to estimate crowd density probability map for crowd counting.The crowd density probability map reflects the frequency of pedestrians appear in an image and the sum of map is equal to the number of people in image.This indirect method relieve the difficulty for crowd counting.However,crowd counting only reflects global density and ignore the information of local density.Crowd density probability map can provide crowd distribution information in 2D image because of perspective distortion.To address this problem,this paper proposes a jointly learned multi-scale fusion and recursive convolutional neural network based framework to estimate the crowd density probability map and crowd perspective map simultaneously.By convolving a perspective adaptive kernel on the crowd density probability map,we can generate a pixel-wise crowd density distribution map in which the pixel value denotes the actual intensity of the crowd at the corresponding location in the real scene.Since the scale size and scale variance of crowd are good cues for estimating both crowd density probability map and perspective map,formulating these two objectives together enables learning a strong feature representation for both tasks.An extension dataset from Shanghaitech crowd dataset B is introduced for the perspective map learning task.Experimental results on 4 datasets demonstrate the effectiveness and reliability of our proposed approach for both crowd density probability map and perspective map.More importantly,we can get the crowd density distribution map which can reflect the density intensity of 3D scene.We apply crowd density distribution map to detect abnormal collection directly or send map sequence to a convolutional neural network for abnormal dispersion detection.Besides,our method can provide extra information about the state of crowd,the location where events occur and quantized abnormal parameters such as velocity and density.Experimental results on 2 datasets demonstrate the effectiveness and reliability of our proposed method.The extra information provided by our method will be more helpful for crowd management.