Font Size: a A A

The Research Of Pedestrian Detection Algorithm Based On Convolution Neural Network

Posted on:2019-11-05Degree:MasterType:Thesis
Country:ChinaCandidate:Y BaiFull Text:PDF
GTID:2428330548954666Subject:Electronic Science and Technology
Abstract/Summary:PDF Full Text Request
At present,the pedestrian detection technology has been extensively studied by researchers,and many related algorithms have been proposed continuously with good results have been obtained.Among them,the pedestrian detection algorithm based on convolutional neural network has gradually become the mainstream of current algorithms.With the expansion of application fields,somespecific high-precision fields set higher requirements for pedestrian detection technology.Therefore,perfecting and improving the pedestrian detection algorithm based on convolutional neural network has important research significance and value.Based on careful analysis of current object recognition,object classification,object detection,pedestrian detection,and crowd counting,this paper proposes a pedestrian detection algorithm based on convolutional neural networks.The related work of this article mainly includes the following aspects:(1)Literature research and theoretical accumulation.Read and analyze relevant documents and algorithms of convolutional neural networks,object recognition,object classification,object detection,pedestrian detection,and crowd counting,summarize the advantages and disadvantages of each algorithm to lay the theoretical foundation to propose the pedestrian detection algorithm based convolutional neural network.(2)Design ofconvolutional neural network model.In order to extract pedestrian characteristics effectively,a convolutional neural network model based on Inception structure was constructed.The new network model is designed for the difference in pedestrian size caused by the distance,height,or other conditions of the shots in the image.The feature is extracted from multi-dimensions of objects in the image,and use the estimated density map as the output of network to retain the features required in the pictureas much as possible.(3)Design of sample labeling algorithm.Targeting the serious obstacles between objects and pedestrians,pedestrians and pedestrians in dense crowds,the head of pedestrian is taken as the target for learning,and the density map is used for sample annotation.The density map represents the location distribution of the pedestrian and the probability of the presence of the pedestrian.(4)Design ofpedestrian positioning algorithm.The method of local maxima is designed to extract the object density of output.After a series of processing operations such as normalization,threshold setting to denoising,local maximum,and removal of approximate object points,a better object positioning result was obtained.(5)Algorithm experimental verification.Different network structures are designed to conduct comparison tests,and comparative tests are performed again using different parameters.Pedestrian detection algorithms are studied from multiple angles in order to expect an optimal structure algorithm.The innovations of this paper are:(1)To solve the seriousobstructed situation between pedestrians and their backgrounds,as well as between pedestrians and pedestrians in the complex context.This article uses the head of pedestrian as the detection object to reduce the characteristics of the object to be measured,in order toaccelerate the operation speed and get better detection results.(2)Design a convolutional neural network model based on Inception structure for feature extraction.This model can extract features of different size objects in the image and estimate the density,to retain more features of pedestrians in the picture.The inadequacies of this article are:(1)Limited by the network model structure,the learning ability of network is limited,and the detection speed is also subject to certain restrictions;(2)Ignoring the recognition of human poses,discarding some identifiable features and producing misjudgments on some non-pedestrian objects,causing some errors.
Keywords/Search Tags:Pedestrian detection, Convolutional neural network, Object recognition, Density map, Pedestrian position
PDF Full Text Request
Related items