Research On Crowd Counting Based On Scale-aware Convolutional Neural Network

With the development of Computer Vision,intelligent video surveillance systems play an important role in public safety.And high-density crowd counting is one of the most important aspects of intelligent video surveillance systems.High-density crowd counting has important applications and research implications in both the commercial sectors and public safety.The current research trend in crowd counting is to achieve accurate counting of high-density crowds in more complex contexts.In recent years,the development of deep learning has driven the development of crowd counting algorithms based on convolutional neural networks,resulting in a significant improvement in the accuracy of crowd counting algorithms.But there are still many challenges in complex contexts: due to the location and angle of the video image capture device,the scale of the people in the captured images varies greatly.This affects the positioning of head areas in the crowd during the counting.Parts of the background in the captured images can often interfere with the counting and be misinterpreted as a crowd.In most convolutional neural networks,the feature extraction process requires the recovery of low-resolution feature maps to high-resolution feature maps,which makes the feature map lose a lot of spatial location information and detail information in the upsampling process.This seriously affects the quality of the density map.The above three problems seriously affect the accuracy of high-density crowd counting.The three problems mentioned above greatly affect the accuracy of crowd counting results,so this paper investigates the relevant network models in the field of crowd counting to address these problems.The main research content and innovations are summarised as follows.(1)In this paper,the Cascaded Parallel Network is designed.The whole network structure consists of three stages,each stage cascad a new branch.These form a network structure.Each branch acquires different scale information through the scale-aware module,and combines with the cross-branch multi-resolution fusion module to interact with each other.For the problem of background interference in crowd counting images,the network is addressed by the proposed filtering layer,which reduces background interference while also achieving the goal of filtering redundant information.To address the low quality of the density maps,the structure of the network adopts the form of cascade.So,the resolution of the feature map in the branch is maintained,and the feature maps are downsampled between the branches by the scale-aware module.so that the feature maps of the first branch contain more detailed information.The feature maps of the third branch contain more semantic information.This enables the feature maps of the three branches to be fused across branches several times to obtain a high-quality density map.(2)To speed up network training and reduce parameter redundancy without losing counting accuracy,the One-Shot Aggregation network with scale-aware is designed.The network consists of a front-end network,scale-aware fusion modules and a channel-wise attention module.The feature is extracted by the front-end network.After that,the input feature maps are fused with the scale-aware fusion module to obtain different scale information.This method solves the problem of perspective distortion in the crowd images.Multiple scale-aware fusion modules are connected to receive input from each layers of the front-end network,which enables the feature map to fuse features from high to low dimensions.This method greatly increases the possibility and diversity of feature aggregation and thus ensures the quality of the output density map.To address the problem of background interference,this network adds a channel-wise attention module before the output layer.The channel-wise attention module suppresses channels with serious background interference.And it enhances the channels that are favorable to counting.This method solves the problem caused by complex and variable backgrounds.