基于级联卷积神经网络和多线索融合的行人检测研究-计算机技术专业论文.docxVIP

优秀毕业论文 精品参考文献资料 ABSTRACTAs ABSTRACT As a key research direction in computer vision，the pedestrian detection technology provides a fundamental support for applications such as object tracking， behavior understanding，intelligent driving system and advanced human—computer interaction system．Due to the non·rigid property of human body，pedestrian detection is vulnerable to complicated background，occlusion，deformation of the body and changes of light conditions，which make it become difficult research task． In this paper，the basic structure of object detection systems and some classic technologies are introduced．And the development of pedestrian detection in the past twenty years and deep learning technologies are discussed．Then the convolutional neural networks are used for detection．The main contributions are follows： (1)A Cascade—CNN detection framework is proposed，in which the first CNN classifies candidate windows roughly and the second one removes false positives elaborately．The mini-batches Stochastic Gradient Descent algorithm is used to train the networks in this framework．Experiment shows that the average miss rate obtained by the proposed Cascade—CNN based on Selective Search algorithm is 48％，which is approximately equal to that obtained by the method based on HOG．When the false positive rate is between 0．1 and 1，our framework outperforms the method based on HOG obviously．In addition，the Cascade—CNN has a simple structure and it runs fast． (2)A multi．cue object proposals extraction algorithm based the Edge Boxes and BING algorithms is proposed．The proposed algorithm call find almost 99．3％ objects in original images and 95．0％objects after the first CNN in 0．223 seconds， which is about a hundred times faster than the same operation using the Selective Search algorithm．The average miss rate obtained by the Cascade·CNN method based multi—cue fusion is 40％，which is 8％less than that obtained by the Cascade-CNN method based the Selective Search algorithm．Moreover，the