Research On Multi-scale Crowd Counting Algorithms With Deep Convolutional Neural Network

With the advent of artificial intelligence era,computer vision has been widely applied in visual systems,such as video surveillance,face recognition and self-driving.Crowd counting is a research hotspot in computer vision and the basis for intelligent video surveillance system.The purpose of crowd counting is to predict the number of pedestrians in still images or videos.In the real scenes,however,the performance of crowd counting algorithm is affected by many factors.Accurate and effective crowd counting is still a challenging work.Among the factors affecting the performance of crowd counting,it is urgent to devise novel method to deal with the scale variations effectively,which is one of the main problems in the research of crowd counting.Owing to the different distances between pedestrians and the camera,the scales of pedestrians are various,which leads to low counting accuracy.Focusing on the problem of scale variations,this thesis based on deep neural network studies the crowd counting models from two aspects: multicolumn convolutional neural network and attention mechanism.The main research contents of this thesis are as follows:(1)To remedy the problem of scale variations in highly congested scenes,a dense crowd counting model based on deep scale purifier network(DSPNet)is proposed.Previous models on crowd counting mainly work on a relatively shallow neural network,which have a limited descriptive ability.And the existing methods have poor generalization ability because of the scarceness of large-scale crowd counting datasets.In order to improve the performance of the model,this thesis proposes an end-to-end model for dense crowd counting.The solution makes use of deconvolution layers and multi-column convolutional neural network to encode multi-scale features.The model through the transfer learning leads to a better generalization ability.Experimental results on dense crowd counting datasets show that the proposed method achieves better counting results and outstanding generalization performance.(2)Aiming at the problem that the spatial and channel features extracted by convolutional neural network have a poor distinguishing ability,a novel crowd counting model combined with dual attention mechanism(SEN-DAB)is proposed.Inspired by the visual attention mechanism,this thesis combines attention mechanism with deep neural network to selectively extract spatial and channel information from original data,which can estimate the number of pedestrians effectively.In addition,a pyramid residual scale enhanced block is presented to learn abundant scale information and further reduce the counting error.Compared with the other crowd counting algorithms,the performance of SEN-DAB achieves significant improvements.Experimental results on public datasets demonstrate the effectiveness of the proposed method.