2017 22 2 CHINESE JOURNAL OF STEREOLOGY AND IMAGE ANALYSIS Vol. 22 No. 2 June 2017 209 1007-1482 2017 02-0209 -0215 DOI 10. 13505 /j. 1007-1482. 2017. 22. 02. 014 RPN 1 2 3 3 1 1 2 3 1. 100090 2. 100090 3. 100090 RPN RPN RPN TP37 A End - to - end monitoring scene pedestrian detection based on improved RPN YANG Dongming 1 2 XU Shibiao 3 MENG Weiliang 3 GE Shuiying 1 YANG Zhen 1 2 ZHANG Xiaopeng 3 1. National Science Library of CAS Beijing 100090 China 2. University of Chinese Academy of Sciences Beijing 100049 China 3. Institute of Automation of CAS Beijing 100090 China Abstract Compared to other scenes pedestrian detection in monitoring scenes gets larger flow of people and higher degree of occlusion. This paper presents an end - to - end detection scheme based on region proposal network RPN. On the one hand we improved RPN by combining with a deep convolution network designed by ourselves to obtain a new deep neural network for pedestrian detection. On the other hand we introduced head - shoulders model to improved detection performance for pedestrian detection at monitoring scenes which improves the detection speed. Eventually we achieved end - to - end detection. The experimental results show that the method effectively improved the detection performance by reducing the missing rate and speeding up the detection. Key words monitor pedestrian detection deep learning RPN network 2017-03 - 30 No. 2012BAH42F02 61332017 61671451 61571046 1993 - E - mail shuiying. ge@ rcsc. cas. cn
210 2017 22 2 0 1 RPN 2 DeepParts 22 part pool 2 3 HOG 12 LBP 13 Harr - like 15 DPM 16 Deformable Parts Model VGG16 X. Wang 14 VGG16 2012 Convolutional neural R - CNN Region of CNN 18 CNNs selective search region proposals SPP - net 19 1 3 Fast - RCNN 20 SPP spatial pyramid pooling ROI region of interest pooling Girshick R Faster R - CNN 20 RPN Region Proposal Network RPN proposal ROI pooling Faster R - CNN RPN 5 RPN anchors R 3. 1 Benenson 7 S Zhang 8 state of art VGG16 23 Oxford Net Andrew Zisserman 2014 ILSVRC localization and classification networks CNNs 17 RPN 6 9-11 CNNs Deep ConvNet RPN Classifier Deep JointDeep 9 ConvNet RPN R - CNN 1
2017 22 2 RPN 211 1 Deep ConvNet RPN Classifier 1 5 ground - truth IoU Intersection - over - 4 Max pooling union 0. 7 0. 7 IoU 3 3 1 ground - truth Relu IoU 0. 3 rpn_bbox_ 512 pred 2 RPN 4 t i t xi t yi t wi t hi K anchor anchor 4K anchor 2 b 3 3 300 512 3 3 L( { p }{ i t }) = 1 i N cls p i p * i + 1 i N reg p * i i L reg t i t * i 2 i i anchor p * i = 3 1 0 t * i ground - true box 1 1 anchor box 1 1. 5 1 2 2 a 128 256 512 K anchor = 3 3 3 RoI pooling RoI pooling N anchor = W Conv5 H Conv5 K anchor 1 ROIs region of interest W Conv5 H Conv5 rpn_cls_score rpn_bbox_pred rpn_cls_score resize P i 0 or 1 crop /wrap K anchor anchor 2K
212 2017 22 2 2 a b RPN 3 3 K anchor anchors 2 K anchor 4 K anchor 3. 2 27% 63% 98% 3 a Chen Q 23 SED - PD 3 a 3 b 4 3 a b CPU Intel R Xeon VOC2007 25 INRIA Person dataset 26 ImageNet R CPU E5-2620 v3 @ 2. 40GHz GPU 12G 0. 001 momentum 0. 9 weight decay TITAN X Pascal Caffe 0. 0005 40 K train_loss /auc - iteatation RPN 4 train_loss - iteatation auc - iteatation 4. 1 0. 352 s /iter INRIA Person dataset 74. 02%
2017 22 2 RPN 213 0. 109 s 9. 17 Faster 1 INRIA Person dataset R - CNN VOC2007 + VOC2012 200 VOC2007 73. 2% 457 5 / 2 INRIA Person dataset 4 4. 2 INRIA Person dataset 1826 3 4. 1 4. 2. 2 5 a 3 a 0. 093 s 10. 75 INRIA Person dataset 4. 2. 1 5 b 5 a 1 a conv5_2_relu5_2 rpn_cls_score rpn_bbox_pred rpn_output b
214 2017 22 2 4. 4. 3 J. 2015 35 6 111-116. DOI 10. 14132 /j. cnki. 1673-5439. 2015. 06. 017. Zeng Min Zhou Yilong. Simulation of pedestrian detection based on deep learning model J. Journal of Nanjing Uni- of Posts and Telccommunications 2015 35 6 1 versity RPN 111-116. DOI 10. 14132/j. cnki. 1673-5439. 2015. 06. 017. 1 1 1 1. 5 1 2 5. J. 2012 40 4 814-820. DOI 10. 3969 /j. issn. 0372-2112. 2012. 04. 031. Su Songzhi Li Shaozi Chen Shuyuan et al. A Survey on 2 pedestrian detection J. Acta Electronica Sinica 2012 VGG16 40 4 814-820. DOI 10. 3969 /j. issn. 0372-2112. RPN 2012. 04. 031. 6 Zeng Xingyu Ouyang Wanli Wang Xiaogang. Multi - stage contextual deep learning for pedestrian detection 5 7 Benenson R Omran M Hosang J et al. Ten years of pe- destrian detection what have we learned C / /ECCV RPN 2014 Workshops 2014 8926 613-627. 8 Zhang S Benenson R Omran M et al. How far are we from solving pedestrian detection C / / in CVPR 2016 1259-1267. 9 Ouyang Wanli Wang Xiaogang. Joint deep learning for pedestrian detection C / / IEEE International Conference on Computer Vision 2013 2056-2063. 10 Luo Ping Tian Yonglong Wang Xiaogang et al. Switchable deep network for pedestrian detection C / /IEEE Conference on Computer Vision and Pattern Recognition 2014 899-906. References 1. D. 2015. 12 Dalal N Triggs B. Histograms of oriented gradients for Wang Bin. Pedestrian detection based on Deep Learning human detection C / /IEEE Computer Society Conference D. Beijing Jiaotong University 2015. on Computer Vision & Pattern Recognition. IEEE Comput- 2. er Society 2005 886-893. D. 2013. Liu Xiaohui. Pedestrian Detection and Tracking in Complex Visual Scenes D. Tianjin University 2013. 13 Ojala T Pietikainen M Maenpaa T. Multiresolution gray - scale and rotation invariant texture classification with local binary patterns C / /PAMI 2002 24 7 971-987. 3. HOG 14 Wang X Han T X Yan S. An HOG - LBP human detector D. 2014. Zhou Kejia. Research on pedestrian detection and tracking with partial occlusion handling C / /IEEE International Conference on Computer Vision. IEEE 2010 32-39. based on HOG and template matching D. Jilin University 2014. C / / IEEE International Conference on Computer Vision 2013 121-128. 11 J. Hosang M Omran R Benenson et al. Taking a deeper look at pedestrians C / /in CVPR 2015 4073-4082. 15 Zhang S Bauckhage C Cremers A B. Informed Haar - Like Features Improve Pedestrian Detection C / /Comput-
2017 22 2 RPN 215 er Vision and Pattern Recognition. 954. IEEE 2014 947-16 Felzenszwalb P Mcallester D Ramanan D. A discriminatively trained multiscale deformable part model C / / IEEE Computer Society Conference on Computer Vision and Pattern Recognition. DBLP 2008 1-8. 17 Krizhevsky A Sutskever I Hinton G E. Imagenet classification with deep convolutional neural networks C / /Advances in Neural Information Processing Systems 2012 1097-1105. 18 Girshick R Donahue J Darrell T et al. Rich feature hierarchies for accurate object detection and semantic segmentation C / /Computer Vision and Pattern Recognition CVPR 2014 IEEE Conference on. IEEE 2014 580-587. 19 He K Zhang X Ren S et al. Spatial pyramid pooling in deep convolutional networics for visual recognition J. ArXiv Preprint ArXiv 1406. 4729 2014. 20 Girshick R. Fast R - CNN C / /IEEE International Conference on Computer Vision IEEE Computer Society 2015 1440-1448. 21 Ren S He K Girshick R et al. Faster R - CNN Towards real - time object detection with region proposal networks J. IEEE Transactions on Pattern Analysis & Machine Intelligence 2016 1-1. 22 Tian Y Luo P Wang X et al. Deep learning strong parts for pedestrian detection C / / IEEE International Conference on Computer Vision IEEE 2015 1904-1912. 23 Simonyan K Zisserman A. Very deep convolutional networks for large - scale image recognition J. In ICLR 2015. 24 Chen Q Jiang W Zhao Y et al. Part - based deep network for pedestrian detection in surveillance videos C / / IEEE 2015 1-4. 25 Everingham M Gool L Williams C K et al. The Pascal visual object classes VOC challenge J. International Journal of Computer Vision 2010 88 2 303-338. 26 Dalal N Triggs B. Histograms of oriented gradients for human detection C / /CVRR 2015 France 2005 1 12 886-893.