Research On The Pedestrian Detection Based On Region Of Convolution Neural Network

Pedestrian detection has been a hotspot in the field of machine vision,and it is widely used in the field of artificial intelligence such as intelligent monitoring,intelligent transportation and intelligent robots.For example,in the field of traffic safety,pedestrian detection technology can be used to predict whether there are pedestrians in the front and near,if found immediately to take emergency braking,which can effectively avoid vehicle collision pedestrians and reduce casualties.Pedestrian detection is different from ordinary object detection and pedestrians are non-rigid objects.In real life,pedestrians wear a wide variety,and the backgrounds are complex,illumination change a lot and occlusion often appeared,which brings a huge challenge to this work.Many of the effective pedestrian detection algorithms have been proposed,and the most representative one is the histogram of oriented gradient(HOG).But in a more complex background environment,the algorithm still has shortcomings.In recent years,deep learning has re-entered the perspective of people,and deep Convolution Neural Network(CNN)has made great success on image and audio,which is the important component of deep learning.On the basis of intensive investigation on the related technologies of pedestrian detection and deep learning especially deep Convolution Neural Network,the main research achievements are summarized as follows:(1)Design a pedestrian detection system based on region of Convolution Neural Network.Artificial designed methods of feature extracting had an imperfect description of pedestrian in the complex background.In this thesis,we propose a pedestrian detection system based on deep learning,adapting a general-purpose convolutional neural network to the task at hand.It can make full use of the advantage of deep convolutional neural network and extract features from the database of pedestrian detection.Because of the layer of the deep learning network architecture is often very deep and more training parameters are needed.We can effectively avoid over-fitting problem when training the network only on the condition that the training data is sufficient.Finally,through the verification of several experiments,compared with the method based on HOG feature,our algorithm can obviously improve the accuracy of pedestrian detection.(2)To solve the problem of low efficiency of generate region proposal in the pedestrian detection system,we use the edge boxes algorithm instead of selective search algorithm to extract region proposal.To acquire high quality region proposal is very important for the pedestrian detection system.However,it takes about 2 seconds to extract region proposal of an image by using the selective search algorithm,which seriously affects the detection efficiency of the whole pedestrian detection system.When we use Edge Boxes algorithm to extract region proposal,although the detection accuracy is not significantly improved,but only takes 0.3 seconds to extract region proposal of an image,greatly improving the detection efficiency of the system.(3)Design a fast pedestrian detection system based on region of Convolution Neural Network.Generally,it is difficult to guarantee the real-time when adapting deep convolutional neural network extract features.In view of this problem,we introduced the RoI Pooling Layer in the network architecture,and we need to extract convolution feature from the original image only once,which can obtain the fixed dimension of the feature vector.Experiments show that the proposed method can greatly improve the detection speed and improve the real-time performance and applicability of the algorithm in the case of ensuring high detection accuracy.