Infrared Pedestrian Detection in Complex Night Scenes
-
摘要: 针对夜间红外图像中行人与背景灰度差异小且存在遮挡等问题,提出了一种夜间复杂场景下的红外行人检测算法。首先利用行人语义融合方法生成对目标全覆盖的显著图,与原图融合得到感兴趣区域,然后构造基于改进的方向梯度直方图特征的两分支分类器,同时提出一种遮挡判别算法,根据分类器模糊分数判断是否遮挡,设计一种头部模板实现最终的行人检测。在LSI远红外行人数据集和自主采集的冬、夏季节夜间行人数据上进行实验,结果表明:在不同环境下,所提出的方法均可快速鲁棒地检测出行人,可较显著地降低漏检率,检测率可达到94.20%。Abstract: An infrared pedestrian detection algorithm is proposed to solve the problem of small differences between pedestrians and backgrounds in gray scale images and the occurrence of occlusion in infrared images at night. First, a significant graph with the full coverage of the target is generated by the pedestrian semantic fusion method, and the region of interest is obtained by combining it with the original graph. Then, a two-branch classifier based on the improved histogram of the gradient feature is constructed. The fuzzy score of the classifier is used to determine the occurrence of occlusion and call the head template for the final detection. Experiments based on the LSI far infrared pedestrian dataset and independent datasets of pedestrians captured at night in winter and summer prove that the proposed method is robust and quick in detecting pedestrians under different environments. It can significantly reduce the rate of missed detection and realize a detection rate of 94.20%.
-
Key words:
- infrared image /
- pedestrian detection /
- saliency /
- complex censes /
- HOG feature
-
表 1 两分支的SVM分类器训练参数
Table 1. SVM classifier training parameters of the two branches
Window size Block size Cell size Step Bin Feature dimension Near target 48×96 16×16 8×8 8 9 1980 Distant target 24×48 8×8 4×4 4 9 1980 表 2 测试数据的基本信息
Table 2. Basic information of test data
Dataset Size/frame Date Location Test-L 2000 - - Test-W1 1578 2018/1/2 Street Test-W2 979 2018/1/3 Campus Test-S1 1123 2019/5/29 Street Test-S2 1089 2019/5/30 Campus 表 3 参数混淆矩阵
Table 3. Parameter confusion matrix
Truth Predicted result Positive example Negative example Positive example TP FN Negative example FP TN 表 4 Wm标记结果
Table 4. Wm marking results
Wm Picture number Pedestrian number Pedestrian mark Labeling rate/% Mark time/s 5 500 774 630 81.4 5.94 10 500 774 636 82.2 8.39 50 500 774 671 86.7 12.70 100 500 774 928 94.1 17.83 150 500 774 751 97.0 23.29 200 500 774 752 97.2 29.01 350 500 774 755 97.6 44.00 500 500 774 755 97.6 56.38 表 5 与基本算法检测结果对比
Table 5. Comparison of detection results with the basicalgorithm
Algorithm ACC/% t/ms HOG 89.27 236.4 TBHOG 86.19 122.8 TBHOG+ROI extraction 91.44 159.0 TBHOG+Occlusion handling 93.23 168.4 TBHOG+ ROI extraction + occlusion handling 94.20 181.6 表 6 与其他算法检测效果对比
Table 6. Comparison of detection effect with other algorithms
Algorithm LSI FIR test-W test-S ACC/% t/ms ACC/% t /ms ACC/% t/ms HOG 89.27 236.4 80.59 289.7 62.77 274.1 LBP 85.21 202.3 77.90 304.5 59.25 294.5 HOG-LBP 93.64 333.7 84.03 375.8 68.17 373.8 ACF 94.05 1008 85.54 1469 64.28 1742 RetinaNet 90.13 1364 82.21 1730 63.09 2297 Proposed 94.20 181.6 89.17 264.7 73.64 285.3 -
[1] XU Z, ZHUANG J, LIU Q, et al. Nighttime FIR pedestrian detection benchmark dataset for ADAS[C]// Proceedings of Pattern Recognition and Computer Vision, 2018: 323-333. [2] TAO Y, FU D, SHU P. Pedestrian tracking for infrared image sequence based on trajectory manifold of spatio-temporal slice[J]. Multimedia Tools and Applications, 2017, 76: 11021-11035. doi: 10.1007/s11042-016-3461-8 [3] 刘洋. 基于LS-DYNA的汽车正面碰撞计算机模拟仿真[D]. 西安: 西华大学, 2011.LIU Yang. Simulation on the Front Impact of Vehicle Based on LS-DYNA[D]. Xi'an: Xihua University, 2011. [4] Dalal N, Triggs B. Histograms of Oriented Gradients for Human Detection[C]// 2005 IEEE Computer Society Conference on Computer Vision and Pattern Recognition (CVPR'05), 2005, 1(1): 886-893. [5] Bosch A, Zisserman A, Munoz X. Representing shape with a spatial pyramid kernel[C]//Acm International Conference on Image & Video Retrieval, 2007: 401-408(doi: https://doi.org/10.1145/1282280.1282340). [6] Sangeetha D, Deepa P. A low-cost and high-performance architecture for robust human detection using histogram of edge oriented gradients[J]. Microprocessors and Microsystems, 2017, 53: 106-119. doi: 10.1016/j.micpro.2017.07.009 [7] ZHENG C H, PEI W J, YAN Q, et al. Pedestrian detection based on gradient and texture feature integration[J]. Neurocomputing, 2017, 228: 71-78. doi: 10.1016/j.neucom.2016.09.085 [8] 朱聪聪, 项志宇. 基于梯度方向和强度直方图的红外行人检测[J]. 计算机工程, 2014, 40(12): 195-198, 204. https://www.cnki.com.cn/Article/CJFDTOTAL-JSJC201412037.htmZHU Congcong, XIANG Zhiyu. Infrared pedestrian detection based on histograms of oriented gradients and intensity[J]. Computer Engineering, 2014, 40(12): 195-198, 204. https://www.cnki.com.cn/Article/CJFDTOTAL-JSJC201412037.htm [9] Itti L, Koch E. A model of saliency-based visual attention for rapid scene analysis[J]. IEEE Trans., 1998, 20(11): 1254-1259 http://dl.acm.org/citation.cfm?id=297870 [10] Radhakrishna A, Sheila H, Francisco E, et al. Frequency-tuned salient region detection[C]//2009 IEEE Conference on Computer Vision and Pattern Recognition, 2009: (doi: 10.1109/CVPR.2009.5206596). [11] HOU X D, ZHANG L Q. Saliency detection: a spectral residual approach[C]//2007 IEEE Conference on Computer Vision and Pattern Recognition, 2007: (doi: 10.1109/CVPR.2007.383267). [12] Alexe B, Deselaers T, Ferrari V. Measuring the objectness of image windows[J]. IEEE Transactions on Software Engineering, 2012, 34(11): 2189-2202. http://newmed.wanfangdata.com.cn/Paper/Detail/PeriodicalPaper_PM22248633 [13] WANG X, HAN T X, YAN S. An HOG-LBP human detector with partial occlusion handling[C]//12th International Conference on Computer Vision of IEEE, 2010: (doi: 10.1109/ICCV.2009.5459207). [14] Javier M N, Vazquez D, Lopez A M, et al. Occlusion handling via random subspace classifiers for human detection[J]. IEEE sTransactions on Cybernetics, 2013, 44(3): V342-354. http://www.ncbi.nlm.nih.gov/pubmed/23757554 [15] Broggi A, Bertozzi M, Fascioli A, et al. Shape-based pedestrian detection[C]//IEEE Intelligent Vehicles Symposium, 2000: (doi: 10.1109/IVS.2000.898344). [16] Brehar R, Vancea C, Nedevschi S. Pedestrian detection in infrared images using aggregated channel features[C]//IEEE International Conference on Intelligent Computer Communication & Processing, 2014: (doi: 10.1109/ICCP.2014.p6936964). [17] LIN T Y, Goyal P, Girshick R, et al. Focal loss for dense object detection[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 99: 2999-3007. doi: 10.1109/ICCV.2017.324 [18] 车凯, 向郑涛, 陈宇峰, 等. 基于改进Fast R-CNN的红外图像行人检测研究[J]. 红外技术, 2018, 40(6): 578-584. http://hwjs.nvir.cn/article/id/hwjs201806010CHE Kai, XIANG Zhengtao, CHEN Yufeng, et al. Research on infrared image pedestrian detection based on improved fast R-CNN[J]. Infrared Technology, 2018, 40(6): 578-584. http://hwjs.nvir.cn/article/id/hwjs201806010