Study On Population Density Estimation Based On Video Image

In recent years,with the increasing number of accidents caused by huge crowds,the study of population density estimation has become a hot topic in video surveillance.At the same time,computer vision technology has been widely applied in everyday life,such as license plate recognition,face detection,fingerprint recognition etc.With the continuous achievement of deep learning,multi-level structure models have been caused attention nowadays.Among them,one of the most representative deep learning models is deep convolution neural network,which extracts deep features of the image through multi-hidden layer network structure,and avoids complex feature design process by feature learning.The different methods of crowd density estimation are studied in this paper,and the structure of convolution neural network is designed applying in crowd density estimation to achieve fast and accurate classification.1)A method combining static and dynamic features is proposed to estimate population density.A new dynamic feature--diffusion rate is extracted with pixel statistics and texture features.Firstly,video image features are extracted,including static characteristics of Gabor filter characteristics,gray co-occurrence matrix,dynamic characteristics of the diffusion rate.After feature extraction,the improved sequential forward method is used to optimize the combination of features.Then the fuzzy wavelet neural network based on support vector machine is used to estimate the population density.The experimental results show that the method has a higher recognition rate than the population density estimation method of the single static feature video image.2)An improved population density algorithm for the convolution neural network is proposed.Convolutional neural network can efficiently and adaptively learn deep characteristics in the feature extraction process,and demonstrates its superiority in depth learning areas,but in the pretreatment oscillation occurs,and the size matching for feature map of convolutional layer and sub sampling layer will affect the calculation speed and time.This paper adopts discrete wavelet transform to replace the sub sampling layer in convolutional neural network,and the weight matrix in the network is adapted to improve the calculation,the phenomenon of oscillation is avoided in pretreatment by adaptive weight.This improves the matching degree of feature map size in convolutional networks,and it is applied in the estimation of crowd density,effectively improves the data correlation and enhances the learning ability of the network,and also improves the accuracy of classification of population density.The experiment shows that the improved network has better learning and classification effect and robustness,and can be used to estimate the population density more accurately and quickly.3)A fast algorithm for population density estimation based on parallel convolution neural network is proposed.Firstly,the image is divided into blocks and used as input to the parallel convolution neural network.Then,different convolution kernels are employed to extract different features based on different regions,and feature fusion is performed.Finally,the classification is carried out and the classification results are obtained.Experiments show that the network has better learning and classification effects and higher operational speed,and it can estimate the crowd density more accurately and quickly.