Crowd Counting Based On Multi-branch Feature Extraction

The population of cities has increased sharply in recent years for the development of economy and urbanization.People tend to gather in public places,leading to frequent safety accidents.Therefore,the crowd counting algorithm used for security research is a hotspot in recent years.However,the performance of the crowd counting algorithm is still facing great challenges due to the problems such as the cluttered background and the different sizes of pedestrians in the crowd image.The work of this paper is as follows:1.For the problem of different-scales pedestrian,this paper proposed a crowd counting algorithm based on stratified close connection.Firstly,the backbone network is divided into three small modules with different depths to extract abundant details characteristics that are beneficial to small-scale information recognition;Based on the residual block,the semantic feature enhancement module of stepped multi-branch structure was designed to extract abundant semantic features that are beneficial to large-scale information recognition;Finally,semantic features and detail features are fused in a tightly connected way to promote the algorithm to have both details features and semantic features,and strengthen the algorithm’s ability of identifying large-scale and small-scale pedestrian information,so as to improve the accuracy of the algorithm.2.For the problem of cluttered background in crowd image,this paper proposed a crowd counting algorithm based on pedestrian information guidance.Firstly,a multi-branch feature extraction module was designed based on dilated convolution with different dilation rates to extract multi-scale pedestrian features;then designed the pedestrian information guide module of refraining background information and highlighting pedestrian information to get significant pedestrian guidance feature maps that guiding the algorithm to focus on multiple-scales pedestrian features,promote the network to assign more weight values to pedestrian information and less weight values to background information,so as to weaken background information,reduce interference of background information and improve algorithm performance.3.The above crowd counting algorithm are compared with the excellent crowd counting algorithm such as Sa CNN,PSDNN and HYGNN to verify their accuracy and robustness;In order to verify the validity and rationality of the designed module,the ablation experiment was carried out on the designed module.The results show that,compared with other algorithms,the MSE of the algorithm proposed in this paper is reduced by more than 3.5% in the part-B of Shanghai Tech data set,the MSE is more than 6.6% in the CUF＿CC＿50 dataset,also have the same good performance in the UCF-QNRF dataset and World-EXPO ’10 dataset,which effectively improves the accuracy of the algorithm.