Crowding Counting Algorithm With Convolutional Neural Network

In recent years,stampede accidents have frequently occurred in large-scale crowd gathering occasions,resulting in a large number of casualties and property losses,and pose a great threat to social public safety.Although the monitoring system has been very popular and can monitor crowd gathering occasions in real time,it still lacks ef-fective measures for crowd management and accident prevention.Accurately obtaining crowd density distributions and total numbers from surveillance pictures and videos can provide effective information for accident prevention.Therefore,in the field of com-puter vision,research on crowd counting has attracted the attention of more and more researchers and has become a subject of high research value.In this dissertation,we study the image crowd counting and propose a image crowd counting algorithm based on convolutional neural network.The algorithm adopts the density map regression method to learn the mapping relationship between image and density map through convolutional neural network.In order to adapt to the scale varia-tion in crowd images and to improve the counting ability for small targets in high-density crowd images,the algorithm uses a convolutional neural network that is divided into two sub-network structures,one is a feature extraction network for processing scale varia-tion issues,and the other is a feature fusion network that enhances the ability of small target counting.At present,multi-column convolutional neural networks and multi-input convolutional neural network structures are commonly used to solve multi-scale issues.However,these two structures have their own deficiencies.In view of the short-comings of these two networks,the feature extraction network proposed in this paper adopts a single-row network structure.By adding some modules with scale-awareness capabilities in a single-row network,the network has the ability to handle scale varia-tion.The feature fusion network combines the feature maps of different scales extracted from the feature extraction network,and the feature maps after the fusion are regressed to generate density map.The network's ability to perceive and count small objects is enhanced by incorporating low-level details in high-level semantic features.This dissertation conducts a series of experiments on ShanghaiTech and UCF_CC₅0 datasets.First,two validation experiments were conducted on the ShanghaiTech dataset.The effectiveness of the feature extraction network and the feature fusion network in re-solving the scale variation issues and the small target counting issues were respectively tested.The experimental results show that the feature extraction network proposed in this dissertation can effectively deal with scale variation issues and the feature fusion networks can improve the algorithm's ability to count small targets.Then,the net-work proposed in this paper are trained and tested respectively at ShanghaiTech and UCF_CC₅0,and the test results are compared with some popular crowd counting al-gorithms.It demonstrates that the proposed algorithm outperforms most current crowd counting algorithms and has excellent counting accuracy and robustness.