Research On Pedestrian Detection Based On Convolutional Neural Network

Pedestrian detection in traffical situation has a critical demand on speed and accuracy.Traditional methods detect fast but lack accuracy,while method based on Convolution Neural Network detects well but can not meet the requirement of speed.And this article aims at achieving improvements in both speed and detection accuracy based on Convolutional Neural Network.First,this article summarizes the basic knowledge of convolutional neural network,including network structure design,loss function,regularization choice,and optimization policy.And then an experiment is conducted to compare the difference of many variants of gradient descent optimization method,offering pratical guidance on optmization method choice.Second,based on “Faster RCNN” method in object detecion,this article realizes pedestrian detection on Caltech pedestrian dataset.Some adjustments are made based on former summaries,including anchors setup in region proposal network and dropout layer's adding in region of interest classifier.Results show that such method and changments can detect pedestrians effectively,compared to traditional methods.Third,according to the influence of convolutional feature map resolution and receptive field on detection accuracy,this article designed a multi-scale region proposal network.By data augmentation of random cropping and resizing,this network achieves better accuracy on medium and big pedestrian detection.Finally,a network compressing method is introduced to boost detect speed on test phase.With the help of singular value decomposition and tucker-2 decomposition,high demensional full connected layer and convolutional layer could be approximated by cascaded low demensional layers.Experiment results show that this method could achieve 4 times acceleration for single layer,1.6 times acceleration for whole network,and 4 times model size compression without apparent accuracy drop.