Pedestrians Detection Based On Region Proposal Network For Autonomous Driving System

With the popularity of self-driving cars,the safety of it has gradually become the focus of attention.When driving on the road,the self-driving car should not only be responsible for the safety of the vehicle users,but also be aware of the surrounding environment in real time,and be responsible for the safety of other road users,too.As one of the road users,whether pedestrians can be detected quickly or not is related to the popularization and promotion of self-driving cars.Therefore,the thesis simplifys the strcture of faster RCNN and uses the VGG-16 network and ZF network to realize pedestrian detection which based on the region proposal network.We regarded the pedestrians as foregound and the non-pedestrian as background to realize pedestrian detection.We analyzed the network structures and resource consumption firstly and reduced the size of model by cutting and modifying the layers that has possessed too many computer resources.Then,using modified networks realized the training of the models that based on region proposal network.Meanwhile,we ploted the convolution weights picture and realized the visualization of features to predict the detection performance.After that,the Caltech Pedestrian Database and images took by myself were used to show the experimental results.Then,we discussed the influences of training parameters and non-maximum suppression.In order to verify the stability of the algorithms,we realized the experiment by using set00^set03 in Caltech Pedestrian Dataset as training set and subset of Daimler Mono Ped.Detection Benchmark Data Set as testing set.Finally,we compared the results with mainstream algorithms in official benchmark.The results show that the miss rate of Ped-VGG16 algorithm has been reduced to 16.9%(FPPI=10^-1)and Ped-ZF for 33.83%in Caltech Pedestrian Detection Benchmark,and the average time has been reduced to 0.14s and 0.05s respectively.The experimental results show that the performance of algorithms is better than most mainstream algorithms,and the Ped-ZF method achieves the real-time detection.