Estimation Of Crowd Density Based On Deep Learning

In recent years,crowd counting has attracted people's attention because of its wide application,such as public security,congestion avoidance,traffic analysis,etc.The purpose of crowd counting is to estimate the number of people in a crowded image or video provided by a surveillance camera.Using computer vision technology to estimate the number of people accurately and steadily is of great significance to public security.At present,the research trend of crowd understanding has developed from crowd counting to displaying population distribution through density map.In general,because complex crowd scenes are affected by various factors,including background noise,occlusion,and scale changes,it is challenging to produce accurate crowd density maps and perform accurate crowd counting in highly crowded and noisy scenes.The existing methods based on neural network often use multiple inputs or multiple network models to extract scale related features,but this inevitably increases the computational burden.In addition,most crowd counting methods only use the features of the last layer for prediction,and ignore the feature information of the middle layer.Therefore,in view of the above problems,this paper has carried out the following work:1.A static image density estimation method based on convolutional neural network is proposed,which is called Multi-scale Dilated Convolution of Convolutional Neural Network(MScCNN).The method uses a single-column network for feature extraction,combined with multi-scale dilated convolution for multi-scale information aggregation,which solves the deficiencies of multi-input and multi-network methods.The multi-scale dilated convolution module utilizes dilated convolution to systematically aggregate multi-scale context information on the premise of not reducing the receptive field,thereby integrating the low-level detail information into high-level semantic features and improving the method's ability to perceive small targets.2.In order to make full use of the feature information of the middle layer,another static image density counting method based on convolutional neural network is designed,called multi-stage feature-based convolutional neural network for crowd counting(MStCNN).The network structure is divided into two parts.One part uses the pre-trained VGG16 basic network with strong transmission capabilities;the other part is a multi-stage feature fusion network,which uses dilated convolution to further densify the features of different maximum pooling layers of VGG16.Figure fitting.This method makes full use of the features of different stages,including low-level semantic features and rich high-level semantic features,and obtains a good fitting effect.3.A method is designed that combines the first two methods,called based on multi-stage and multi-scale features convolutional neural networks for crowd counting(MStScCNN).Considering the network structure of the first two networks for the middle layer features and the features of different scales respectively,the method combines the above two networks to obtain richer context information,and fully uses the features of different stages and different scales.At last,this paper demonstrates the proposed network structure in ShanghaiTech dataset,UCF?CC?50 dataset and worldexpo'10 dataset,and compares the results with numbers of mainstream crowd counting methods,proves that our method surpasses current state-of-the-art methods and has excellent counting accuracy and robustness.