E-mail Alert
RSS
| Citation: |
Peng Zhirong, Zhao Meirong, Yang Weiming, et al. Improved multispectral aggregate channel for pedestrian detection[J]. Opto-Electronic Engineering, 2017, 44(9): 882-887. doi: 10.3969/j.issn.1003-501X.2017.09.004
|
Improved multispectral aggregate channel for pedestrian detection
-
Abstract
To solve the problem of low pedestrian detection accuracy in a single band due to variability and complexity of external environment, an improved pedestrian detection algorithm based on multispectral aggregate channel feature is proposed. The aggregate channel features of visible images and infrared images are extracted, respectively. The pixel contrast rule is changed and the results are compared with the adaptive threshold. The improved central symmetric local binary pattern feature is added to the feature channels. Different filter banks are designed to filter the multispectral aggregate channel features. The classifier is trained to realize the multispectral pedestrian detection. Experiments show that the improved local binary pattern feature can describe the symmetry of pedestrians of infrared images better and the intermediate filter layer enriches the candidate feature pool. The algorithm can effectively detect pedestrians in various scenes and improve the pedestrian detection accuracy. Compared with the previous multispectral aggregate channel detection work, the algorithm reduces the log-average miss rate.-
Keywords:
- pedestrian detection /
- multispectral /
- local binary pattern /
- aggregate channel /
- filter banks
-
-
References
[1] Zhang S S, Benenson R, Omran M, et al. How far are we from solving pedestrian detection?[C]// Proceedings of IEEEConference on Computer Vision and Pattern Recognition, 2016: 1259–1267. https://www.researchgate.net/publication/301854437_How_Far_are_We_from_Solving_Pedestrian_Detection [2] 黄富瑜, 沈学举, 刘旭敏, 等.基于空时域融合处理检测超大视场红外目标[J].光学精密工程, 2015, 23(8): 2328–2338. Huang Fuyu, Shen Xueju, Liu Xumin, et al. Detection of super wide-field infrared target based on spatial-temporal fusion processing[J]. Optics and Precision Engineering, 2015, 23(8): 2328–2338. [3] Davis J, Sharma V. Background-subtraction using contour-based fusion of thermal and visible imagery[J]. Computer Vision and Image Understanding, 2007, 106(2): 162–182. [4] Castillo J C, Serrano-Cuerda J, Femández-Caballero A, et al. Hierarchical Architecture for Robust People Detection by Fusion of Infrared and Visible Video[M]// Novais P, Camacho D, Analide C, et al. Intelligent Distributed Computing IX. Studies in Computational Intelligence. Cham: Springer, 2016: 343–351. [5] Yuan Y, Lu X Q, Chen X. Multi-spectral pedestrian detection[J]. Signal Processing, 2015, 110(5): 94–100. [6] Hwang S, Park J, Kim N, et al. Multispectral pedestrian detection: Benchmark dataset and baseline[C]// Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2015: 1037–1045. [7] Dollar P, Appel R, Belongie S, et al. Fast feature pyramids for object detection[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2014, 36(8): 1532–1545. doi: 10.1109/TPAMI.2014.2300479 [8] Heikkila M, Pietikainen M, Schmid C. Description of interest regions with center-symmetric local binary patterns[C]// Kalra P K, Peleg S. Computer Vision. Graphics and Image Processing. Berlin, Heidelberg: Springer 2006: 58–69. [9] 刘琳, 耿俊梅, 顾国华, 等.轮廓特征与神经网络相结合的行人检测[J].光电工程, 2014, 41(7): 50–56. Liu Lin, Geng Junmei, Gu Guohua, et al. Human detection based on contour features and neural networks[J]. Opto-Electronic Engineering, 2014, 41(7): 50–56. [10] 谷灵康, 周鸣争.多特征融合下的快速行人检测[J].光电工程, 2015, 42(9): 14–20. Gu Lingkang, Zhou Mingzheng. Fast pedestrian detection based on multi-feature fusion[J]. Opto-Electronic Engineering, 2015, 42(9): 14–20. [11] Nam W, Dollar P, Han J. Local decorrelation for improved pedestrian detection[C]// Proceedings of the 27th International Conference on Neural Information Processing Systems, 2014: 424–432. [12] Zhang S S, Benenson R, Schiele B. Filtered channel features for pedestrian detection[C]// Proceedings of 2015 IEEE Conference on Computer Vision and Pattern Recognition, 2015: 1751–1760. [13] 金志刚, 赵明昕, 张瑞, 等.基于双目视觉的聚合积分通道行人检测优化算法[J].天津大学学报:自然科学与工程技术版, 2016, 49(12): 1225–1230. Jin Zhigang, Zhao Mingxin, Zhang Rui, et al. Optimized aggregated channel features pedestrian detection algorithm based on binocular vision[J]. Journal of Tianjin University: Science and Technology, 2016, 49(12): 1225–1230. [14] Dollar P, Wojek C, Schiele B, et al. Pedestrian detection: An evaluation of the state of the art[J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2012, 34(4): 743–761. doi: 10.1109/TPAMI.2011.155 -
Overview
Pedestrian detection is the principal technique for various applications, such as surveillance, tracking system and autonomous driving. Although the topic has been intensively investigated and significant improvement has been achieved in recent years, pedestrian detection is still a challenging task, limited by occluded appearances, cluttered backgrounds, and low image resolution. Besides, since most of recent researches focus on the detection of pedestrians in visible spectrum images, they are very likely to be stuck with images captured at night or bad lighting. However, ambient lighting has less effect on thermal imaging. Thermal images usually present clear silhouettes of human, but lose fine visual details of pedestrian, which can be captured by visual cameras. To overcome the drawbacks of visible images, it's helpful to fuse the information of visible images and long wave length infrared images. Aggregate channel feature is an easy but useful way to detect pedestrian. However, it only uses the information of visible spectrum images. For the above reasons, an improved pedestrian detection algorithm based on multispectral aggregate channel features is proposed. First, the aggregate channel features of the visible image and the infrared image are extracted, respectively. Specifically, the channel features extracted from the visible images include three LUV color channel features, one normalized gradient magnitude channel feature, and six histogram of oriented gradients channel (HOG) features. The channel features extracted from infrared images include one brightness channel feature and nine HOG features. All the channel features make up the aggregate multispectral channel features. Then, to use the symmetry information of pedestrian in infrared images, the improved central symmetric local binary pattern is proposed. The improved pattern feature is achieved by changing the pixel contrast rule and comparing the contrast result with the adaptive threshold. The improved central symmetric local binary pattern feature is added to feature channels to get the aggregate multispectral channel features. Next, to learn more local features and observe the effect of filters, different filter banks are designed to filter the aggregate multispectral channel features. Finally, the real adaptive boosting learning method is used to train the classifier to realize the multispectral pedestrian detection. Experiments show that the improved local binary pattern feature can better describe the symmetry of pedestrians of infrared images and the intermediate filter layer enriches the candidate feature pool. The algorithm makes use of the complementary information provided by color and thermal images, which can effectively detect pedestrians in various scenes and improve pedestrian detection accuracy. Compared with the previous multispectral aggregate channel detection work, the algorithm reduces the log-average miss rate.
-
Access History
Export File
Citation
Peng Zhirong, Zhao Meirong, Yang Weiming, et al. Improved multispectral aggregate channel for pedestrian detection[J]. Opto-Electronic Engineering, 2017, 44(9): 882-887. doi: 10.3969/j.issn.1003-501X.2017.09.004
Format
Content
DownLoad:
-
Figure 1.
Flow chart of improved multispectral aggregate channel feature for pedestrian detection.
-
Figure 2.
LBP and CS-LBP features for a neighborhood of eight pixels.
-
Figure 3.
Illustration of different filter banks. (a) LDCF filter bank. (b) Random filter bank. (c) Checkerboard filter bank.
-
Figure 4.
Results of pedestrian detection.
-
Figure 5.
Detection results of different scenes. (a) Detection results of a visual images trained by visual band. (b) Detection results of visual images trained by MMACF. (c) Detection results of infrared images trained by MMACF.
