Research On Human Body Detection Method Based On Convolutional Neural Network

Human body detection is a very important research topic in the field of computer vision.Human body detection is a process of distinguishing whether there are human bodes targets in the image.There is high application value in many high-tech fields such as artificial intelligence,intelligent video surveillance,intelligent vehicle assist,intelligent human-computer interaction system and so on.This thesis adopts a human body detection method based on convolutional neural network.Convolutional neural network is an emerging pattern recognition method that combines the deep learning theory into artificial neural networks.The traditional method of human body detection usually extracts the feature firstly,then input the characteristic description to a classifier for training and learning.However,the process of manually extracting features is complex and depends on specific tasks,so the researchers are required high academic level and experience.Convolutional neural network does not need to extract the artificial characteristics of the image in advance,but simulate the human visual nervous system and process the original image and identify layer by layer.This method greatly reduces the training parameters of the network and has good robustness to a certain degree of deformation in image because of partial connection,weight sharing and down sampling.In this thesis,the main work is as follows:Firstly,based on the deep study of convolutional neural network theory,the structure and algorithm of convolutional neural network are analyzed in this thesis,and several network models are designed according to the different network parameters such as convolution kernel,network depth and feature dimension.Secondly,the INRIA database is used as the training sample to experiment with the above network models.By comparing the recognition effect,the influence of the relevant parameters on the network is analyzed,and the optimal network model is selected.This network's training starts to converge when the 10th epochs,and the recognition rate reaches 95.56%.Thirdly,in view of that methods of deep learning usually needs large sample set,this thesis adopts an improved algorithm based on the small sample set,introducing the random dropout.It sets a part of the node to zero randomly,and keeps its weight not updated.Experiments are performed on the improved model in the INRIA subset and on the self-built data set.Experiments show that this method can improve the recognition rate in the case of small sample,effectively alleviate the phenomenon of over-fitting.