Dense Group Behavior Recognition In Video Based On Convolutional Neural Networks

Group behavior recognition in video is a challenging task in computer vision.Due to the problems of scene change,complex crowd distribution,perspective effect and so on,it brings many research difficulties to crowd density analysis,crowd detection and group behavior analysis.Convolutional neural networks provides a reliable solution to the above difficulties.This thesis aims at crowd density analysis and behavior recognition.Firstly,a crowd density detection method based on deeply separated dilated convolutional neural networks is designed,and then a crowd detection and location algorithm based on density map regression guidance is designed,and finally,a crowd behavior analysis method based on density classification is designed to complete the video crowd density analysis and behavior detection.The innovative work of this thesis is as follows:(1)Proposing a depthwise separable dilated convolution networks model for crowd density analysis and crowd counting in highly crowded scenes.This model can more effectively extract the multi-scale features of the image,and use the dilated convolution neural networks to expand the receptive field without increasing the parameters,and add a separation layer and use different dilated rates to overcome the grid effect of dilated convolution networks.It can improve the calculation efficiency of the model and generates a high-quality crowd density map.The model is evaluated on the Shanghai Tech?A / B data set and the UCF-CC-50 data set.The results of experiments show that the prediction error of the model is small,the predicted population density probability map is of high quality and similar to the real density image distribution.(2)Depth adaptive Gaussian function is used to generate crowd density map based on depth image.Depth image is use to cover the crowd density map and input it into the Retina Mask target detection network.Detecting different-sized heads through different decoding layers in the target detection network,and estimating the head size through the depth information of the human head to initialize the anchor point information and estimate the size of the head detection bounding box to realize the detection and positioning of the human head.This method improves the classification rate of the target detection network and effectively solves the problem of missing detection of small-sized heads in crowded scenes.The method is verified on the MICC data set and the Shanghai RGBD data set.The results of experiments show that the method can detect smaller size heads,which effectively improves the detection accuracy and detection rate.(3)Dividing the video crowd density level.Building a spatio-temporal correlation model based on the video acquisition time information and the spatial information of the person's movement in a scene with low crowd density to analyze the group human behavior.For scenes with high crowd density,the crowd's abnormal motion behavior is detected and warned based on the crowd density map.Detecting the number and location of abnormally crowded people,and detecting the number,location,and speed of abnormally dispersed people.The validity of the behavioral semantic extraction method is verified on the Volleyball dataset,and the abnormal motion behavior of the crowd is detected on the PETS 2009 dataset.The experimental results show that the method has high detection accuracy and can locate the location of the abnormality.