Crowd Status Analysis Based On Convolutional Neural Network

The analysis of crowd status has become an important research topic in the field of intelligent image and video analysis.High-precision estimation of crowd amount and detection of movement state are the basis of population state analysis.It can not only help staff optimize management by counting the number of people in the current scene,but also estimate the movement direction and velocity distribution of the crowd to realize the prediction of the trend of human flow.More importantly,it has extremely important value and wide application prospects in preventing overcrowding and detecting potential security issues.For the population statistics,the method of feature extraction is used to generate the feature density map.The density not only contains the population quantity information,but also contains the spatial information of the population distribution.Compared with the traditional method of detecting a single individual to count the population,this method has a wide application scenario,and can not only estimate the number of people in high-crowded and large-number scenarios.At the same time,it can also complete the statistics of population location distribution information.The analysis of crowd-status is also a challenging research.The existing methods still have great limitations in the extraction and processing of application scenarios and features.Therefore,this thesis proposes a method of crowd-state analysis based on convolutional neural network.The method is based on the deep learning algorithm and overcomes the problem of single feature extraction and poor generalization ability in traditional image processing algorithms.In view of the judgment of crowd movement direction in video data,a method of motion optical flow extraction based on CNN is designed.Compared with traditional optical flow extraction method,this method not only has high accuracy and robustness,but also has incomparable advantages in the detection and analysis process of crowd abnormal state with the method of crowd density estimation designed by us.In order to further reduce the constraints of the scene,a crowd density estimation method based on image data is proposed for the image data acquired by the mobile camera with constantly adjusting and changing background.The method takes a single image as input,and generates density map by convolution,pooling and deconvolution.In the analysis of crowd abnormal state,a quantitative method of crowd crowding degree is proposed by fully mining the information of population number and spatial distribution on crowd density map,which provides a basis for judging crowd abnormality in a single image and completes the quantitative evaluation of crowd crowding degree in different scenarios.In view of the image data acquired by more common fixed monitoring devices,more abundant information can be used.Therefore,this thesis proposes a multi-task crowd number and motion optical flow estimation method.The method combines the number estimation task with the motion optical flow extraction task to extract features.The network inputs two consecutive frames of images.The two tasks are shared in the convolution layer and the pooling layer,and the parameters of the two tasks are separated in the deconvolution layer.This design improves the generalization of the whole network,improves the efficiency and reduces the parameters of the model.More importantly,in the judgment of abnormal state of the crowd,a multi-state anomaly detection method is realized by using the detection results comprehensively.Compared with the existing methods,this method is more flexible,practical and accurate.In this thesis,a method of estimating crowd density and extracting crowd optical flow is designed.By combining these methods,the corresponding crowd number information,population spatial distribution information,movement direction and velocity distribution information are generated.According to these information,the quantitative methods of crowding degree and abnormal state of crowd based on multi-state variables are proposed respectively.Finally,the reliability of these methods is verified by experiments.