Crowd Density Estimation In Public Places Based On Deep Learning

With the increase of human social activities,the aggregation of people in every scene has become an essential phenomenon in social activities,such as bus stations,stadiums and subway,which increases the probability of the accident.How to effectively estimate dense scene density has become the focus of many researchers.To solve these problems,many researchers have proposed a variety of algorithms for crowd density estimation.For example,an analysis method based on individual features,a statistical method based on pixels,or a texture based analysis method.The traditional algorithm has obvious advantages in computing speed,but it is especially inadequate in density estimation accuracy,especially when the scene is seriously occluded,and the traditional algorithm is not robust to density estimation of high density scene.Deep learning in the crowd density estimation of catch up from behind.Because of its good robustness and high recognition precision,it has been widely popularized.In this paper,we use depth learning to estimate density,and divide it into two scenarios:low density and high density according to the density level.The reason for the division is to adapt to the transformation of the scene,and the density estimation based on pedestrian detection at low and medium density has better experimental results.The characteristics of the inaccuracy and the poor robustness of the traditional method are effectively solved.Meanwhile,in order to enable this method to be applied in embedded development and wider industrial field,we use network compression method to compress the depth learning model,and achieve the purpose of compressing network without affecting the accuracy.1.For the middle and low density crowd,the method of person detection is used to count the number of the static images,a cascade multitask neural network is proposed.The advantage of the algorithm is the use of the full convolution network,which can be input for an image of any size.It has obvious effect on solving small target pedestrians.At the same time,the idea of cascading networks is used,and the detection accuracy has been greatly improved.Moreover,we use the multi task deep neural network to classify foreground and background.At the same time,we use frame regression to make pedestrian border correction,which can eventually return to pedestrian area accurately,and further improve the accuracy of pedestrian detection.2.For the high density crowd,in this paper,an open data set is used to study the algorithm.A multi column effective convolution kernel network(MECK)is proposed to analyze the scene density.We optimized the loss function,and the experiment analyzed the best multiple of loss function,which made.the network develop in a more robust direction in training.At the same time,we amplify the dataset.On the basis of the original data set,we transform the image to different degrees by different algorithms.The advantage is that the network can adapt to different scenarios and increase the robustness of the network.Furthermore,we optimize the training of the network,and solve the problem of non robust and over fitting due to the lack of data sets.The algorithm is superior to the traditional algorithm in accuracy and robustness,and has some reference for population density estimation.3.In this paper,a custom network compression model is used to compress multiple classes of effective convolution networks to achieve the purpose of compressing the network without affecting the accuracy.We proposes how to compress the network effectively for the weight of the network,that is,how to ensure the accuracy of the deep network compression under the condition of constant accuracy.It is the starting point of this chapter to ensure that the network depth is constant and the weight of the network is minimum.In deep learning,it is difficult to apply the population density estimation algorithm to the embedded end because of the complexity of the network and the large amount of computation.This paper attempts to concentrate the network model,simplify the network structure.The depth of the network is compressed successfully.A great improvement has been made on the basis of the original network,to a certain extent,it solves the embedded application of depth learning in the population density estimation algorithm.