A New Pedestrian Detection Method With Convolutional Neural Network And SVM

Pedestrian detection is an algorithm which detects whether the target picture or video contains pedestrians.If pedestrians are found,they should be marked with rectangular boxes.Pedestrian detection and multi-object detection have become important topics in current computer vision research.At the same time,they have wide application prospect in many areas of real life,including: intelligent transportation,autonomous driving and so on.In recent years,with the development of artificial intelligence,neural network models have been widely applied to the field of computer vision and they have achieved top performance.Therefore,this work proposed a new method that utilizes deep convolutional neural network and SVM to implement pedestrian detection.The innovation of this thesis includes:Firstly,we collect data sets,design network structures,and train a neural network(PedConvNet)to classify pedestrians.The proposed network finally achieved accuracy of 91% on the test set.Secondly,we construct the region proposal network which generates regions where pedestrians may appear in the picture.In order to design a more accurate network,we selected the AlexNet-based network,PedConvNet-based network and VggNet-based network to run comparison experiments.Experimental results finally show that the VggNet-based network generated more accurate region proposal boxes.Moreover,to achieve better detection performance,we extract convolutional features with more distinguishable ability for the SVM classifier.All of the convolutional features computed from Vgg Net and Ped ConvNet are discussed and the results proved that the convolutional features obtained from PedConvNet have better performance.In addition,the results also reveal that the deeper convolutional features usually have more distinguishable ability than the shallow ones and the visualization of the convolutional features verifies this fact as well.Finally,experimental results show that our proposed model achieves 9.1% and 1.26% miss rate on the“reasonable” and “near” setting of the Caltech data set.Comparing to up-to-date pedestrian detection approaches,the high performance of our method shows its feasibility.