Pedestrian Detection Based On Convolutional Neural Network

With the continuous development of computer vision,pedestrian detection technology has attracted more and more attention,and it's application is more widespread.Recently,the performance of computer hardware is improving,and the video data generated in daily life are increasing rapidly.These provide a good opportunity for the research of pedestrian detection based on massive data and machine learning.However,pedestrian detection in the actual scene still faces the challenges of many adverse factors,such as background interference and occlusion.Therefore,how to create a real-time pedestrian detection system with high accuracy is a problem that must be solved.The feature representation ability is a key factor for pedestrian detection algorithm.The lack of representation ability makes classifier fail to discriminate foreground and background.When facing optional shooting angle and multi-pose target,the lack of the robustness will also cause the detection failure.Moreover,Convolutional Neural Network(CNN)plays a significant role in image feature extraction,object classification and recognition.A large number of advanced detection methods based on CNN are emerging,which promotes the development of pedestrian detection technology.This paper improves pedestrian detection algorithm in many respects based on intensive analysis,the main contents are as follows:(1)The pivotal process of pedestrian detection are summarized,such as candidate regions generation,pedestrian features construction,learning and classification methods,and so on.Advantages and disadvantages of typical technologies are analyzed in detail.In order to solve the problem of insufficient robustness of the feature descriptor,we propose a novel method of combining multi-level features based on filtered channel features and CNN.Then the decision trees are trained on abundant features.Finally the strong classifier is learned by Adaboost strategy for pedestrian detection,therefore the accuracy of pedestrian detectionis improved.(2)Several pedestrian detection methods based on CNN and the convolution network construction are illustrated.Then we found that the Convolutional Neural Network plays an important role in the pedestrian detection.Furthermore,Non Maximal Suppression combined with hard thresholding is the most common post-process method in pedestrian detection,while it is easy to cause false negative and false positive.In view of this,we present a novel method based on Faster RCNN and optimized post-processing.In this study,a series of candidate proposals are further selected according to feature similarity,and therefore the false positive and false negative are reduced.We analyze and verify the important role of robust feature and optimized post-processing in pedestrian detection.Experimental results show that proposed pedestrian detection methods have higher accuracy.