Research And Application Of Pedestrian Detection Based On Deep Learning

Focusing on challenges for pedestrian detection which typically include large variations of the appearance of human body,harsh environment and occlusion,this dissertation aims to improve the performance of pedestrian detection technology based on deep learning algorithm,while analyzing applicable scenarios for these detection techniques as well.Above all,the present thesis studies the pedestrian detection method based on Faster R-CNN deep learning,transforming pedestrian detection into classification of pedestrian region proposals.Firstly,region proposals of a pedestrian are extracted by a Region Proposal Network(RPN).Then,for each object proposal,a region of interest(RoI)pooling layer extracts a fixed-length feature vector from the feature map.Each feature vector is finally fed into a sequence of fully connected(fc)layers to produce softmax probability of the pedestrian and pedestrian position.The detection performance is improved by adjusting the network and training parameters and then retraining the detector end to end.The average miss rate of the trained Faster R-CNN pedestrian detector evaluated on the INRIA dataset is 9%,and 19%evaluated on the Caltech dataset,which are both lower than that of all traditional methods.Compared to HOG method,they are respectively 37%and 49%lower.This verifies the effectiveness and practicality of Faster R-CNN deep learning algorithm in pedestrian detection.Then,the pedestrian detection based on R-FCN deep learning method is studied.The R-FCN uses the ResNet as the basic network.In addition to extracting the region proposals of pedestrian,the method also extracts the features that are sensitive to relative positions of a pedestrian,which makes the detector well adaptive to the translation-variant of a pedestrian during the detection.In the scenario with severe occlusion,the average miss rate of R-FCN method is lower than that of all traditional methods and two deep learning methods(RPN + BF and F-DNN),44 percentage point lower than HOG and 22 percentage point lower than RPN + BF,which confirmed that under the circumstance with severe occlusion to pedestrians,R-FCN method still has a good detection performance.Last but not least,the thesis explores the multi-scale pedestrian detection based on SSD in real life scenarios.The SSD detection model developed further reduces the average miss rate and improves the detection speed on the basis of Faster R-CNN.In INRIA dataset,average miss rate of the SSD detection model is 8%,1 percentage point higher than that of RPN + BF and F-DNN Methods,but the detection speed is greatly improved.The results prove that SSD pedestrian detector is of high practicality in real life scenes.Furthermore,comprehensive performances of the retrained Faster R-CNN,R-FCN and SSD detectors are also analyzed.