Pedestrian Detection Based On Convolutional Neural Networks

Pedestrian detection is a technology which determines whether or not there exist people in a given video sequence or static pictures and predicts all the bounding-boxes of pedestrians for them by the computer.If there exist people,it also needs to segment every people from the background and precisely predict the locations for them.Today,pedestrian detection technology has been widely used in the fields of intelligent video surveillance,intelligent robots,intelligent transportation,driver assistance system and so on,and it is also the basic work of pedestrian tracking and people re-identification.Although many great breakthroughs have been made on this technology,there are still many problems to be solved,which are produced by the diversity of pedestrian body poses,the appearance variety and complicated backgrounds.There are two key problems in pedestrian detection task.The first one is to extract robust features.Traditional methods use hand-craft features such as HOG(Histogram of Oriented Gradient)features etc.However,these features only take advantage of low-level or mid-level information in the image,while neglect the higher semantic information.Locating the potential windows which may contain pedestrians is also important.Traditional methods are time consuming and may produce many redundancy windows which may reduce the detection performance.This dissertation focuses on the above two problems and the main research contents are as follows:(1)This dissertation leverages deep learning method,and design a pedestrian detection framework based on Fast R-CNN method.This method can extract robust features which can be a good solution to the problem of semantic gap.(2)In order to extract high-quality region proposals to improve the detection performance,we employ the EdgeBoxes algorithm.It can extract higher quality and low redundancy of candidate windows.(3)In order to reduce the training time and prevent the gradients that explode when training the network,we add the batch normalization layer between the convolutional layer and activation function layer of the Fast R-CNN.Experiments show that the proposed method achieves satisfactory performance on the INRIA and ETH datasets.