Detection of Vehicle Crews Based on Modified Faster R-CNN
-
摘要: 针对现有以雷达技术和红外热成像技术为代表的HOV(High occupancy vehiclelane)车道车辆乘员数量检测方法可靠性差、准确率低等问题,提出一种基于多光谱红外图像与改进Faster R-CNN(Region-Convolutional Neural Networks)的车辆乘员数量检测方法。通过多光谱红外成像系统获得汽车内部空间图像,结合Faster R-CNN深度学习算法实现乘员数量检测,通过采用全卷积网络结构、多尺度特征预测、使用ROI-Align代替ROI-Pooling等方式增强网络的泛化能力。通过对样据进行K-means聚类得到目标框长宽几何比例先验分布,提高区域生成(region proposal network,RPN)网络训练速度和位置回归准确性。测试结果表明,获得的汽车内部空间图像较为清晰,算法可以实现对乘员数量的检测。经过改进,网络的泛化能力得到增强,单乘员检测的准确率达到88.6%,相比于改进前提高了13.8%,能够满足行业规定大于80%的要求。
-
关键词:
- 多光谱红外图像 /
- Faster-RCNN /
- 全卷积 /
- K-means聚类 /
- ROI-Align
Abstract: Existing methods for detecting the number of vehicle occupants in a high-occupancy vehicle (HOV) lane, using radar and infrared thermal imaging technology, exhibit low reliability and low accuracy. To address these limitations, a method for detecting the number of vehicle occupants based on multispectral infrared imaging and an improved Faster regions with convolutional neural networks (R-CNN) algorithm is proposed. The vehicle interior space image is obtained using a multispectral infrared imaging system, and the number of passengers is detected by a Faster R-CNN deep learning algorithm. The generalization ability of the network is enhanced using the full convolution network structure and multiscale feature prediction, and ROI-Align is used instead of ROI-Pooling. Through K-means clustering, the prior distribution of the geometric proportion of the length and width of the target frame is obtained, which improves the training speed and the accuracy of position regression of the region proposal network (RPN). The test results showed that the interior space image was clear, and the algorithm could detect the number of passengers. After its improvement, the generalization ability of the network was enhanced, and the accuracy of single occupant detection reached 88.6%, which was 13.8% higher than before its improvement. This meets the requirements of more than 80% of industry regulations.-
Key words:
- multispectral infrared image /
- faster-RCNN /
- full convolution /
- K-means clustering /
- ROI-Align
-
表 1 RPN改进前后对比
Table 1. RPN Comparison before and after improvement
Method Regions AP/% Times/ms Ori_RPN 2000 69.00 235 Our_RPN1 2000 74.43 240 Our_RPN2 2000 75.52 241 Our_RPN1+Net-Improved
Our_RPN2+Net-Improved2000
200074.91
76.17237
237表 2 不同算法检测效果对比
Table 2. Comparison of detection results of different algorithms
Network Accuracy/% Fps 1 2 3 4 5 Faster R-cnn 74.8 72.6 69.8 65.4 58.8 17 Ours 88.6 86.2 83.0 78.8 73.2 15 YOLOv3 68.6 65.2 61.8 57.4 51.2 30 Mask R-CNN 82.4 78.8 76.2 72.4 66.8 5 -
[1] 曾炎盛. HOV车道实施效果评价体系研究——以深圳市为例[J].交通科技与经济, 2019, 21(5): 39-43. http://www.cnki.com.cn/Article/CJFDTotal-KJJJ201905007.htmZENG Yansheng. Study on the evaluation system of the implementation effect of HOV Lane -- Taking Shenzhen as an example[J]. Transportation Technology and Economy, 2019, 21(5): 39-43. http://www.cnki.com.cn/Article/CJFDTotal-KJJJ201905007.htm [2] 陈鲁峰, 叶鹏飞, 余佳欢, 等. HOV在城市交通管理中的应用可行性分析[J].交通企业管理, 2016, 31(4): 38-41. http://www.cnki.com.cn/Article/CJFDTotal-JTQG201604016.htmCHEN Lufeng, YE Pengfei, YU Jiahuan, et al. Feasibility analysis of HOV application in urban traffic management[J]. Traffic Enterprise Management, 2016, 31(4): 38-41. http://www.cnki.com.cn/Article/CJFDTotal-JTQG201604016.htm [3] 程建梅. HOV专用车道管理思路探索[J].行业管理, 2018(4): 34-35. http://www.cnki.com.cn/Article/CJFDTotal-JTQG201804017.htmCHENG Jianmei. Exploration of management ideas of HOV special lane[J]. Industry Management, 2018(4): 34-35. http://www.cnki.com.cn/Article/CJFDTotal-JTQG201804017.htm [4] Fadel Adib, Zachary Kabelac, Dina Katabi. Multi-person Lcalization via RF Body Reflections[C]//Proceedings of the 12th USENIX Conference on Networked Systems Design and Implementation (NSDI'15), 2015: 279-292. [5] 马也, 常青, 胡谋法.复杂背景下红外人体目标检测算法研究[J].红外技术, 2017, 39(11): 1038-1053. http://www.cnki.com.cn/Article/CJFDTOTAL-HWJS201711012.htmMA Ye, CHANG Qing, HU Moufa. Research on infrared human target detection algorithm in complex background[J]. Infrared Technology, 2017, 39(11): 1038-1053. http://www.cnki.com.cn/Article/CJFDTOTAL-HWJS201711012.htm [6] 马翰飞, 范海震, 李强, 等.基于多光谱融合图像的飞机导航系统设计[J].电子设计工程, 2019, 27(24): 161-166. http://www.cnki.com.cn/Article/CJFDTotal-GWDZ201924035.htmMA Hanfei, FAN Haizhen, LI Qiang, et al. Design of aircraft navigation system based on multispectral fusion image[J]. Electronic Design Engineering, 2019, 27(24): 161-166. http://www.cnki.com.cn/Article/CJFDTotal-GWDZ201924035.htm [7] 李雪欣, 马保东, 张嵩, 等.融合多光谱与SAR影像的地物分类研究[J].测绘与空间地理信息, 2019, 42(12): 55-58. http://www.cnki.com.cn/Article/CJFDTotal-DBCH201912017.htmLI Xuexin, MA Baodong, ZHANG Song, et al. Study on the classification of ground features using multispectral and SAR images[J]. Mapping and Spatial Geographic Information, 2019, 42(12): 55-58. http://www.cnki.com.cn/Article/CJFDTotal-DBCH201912017.htm [8] 赵庆展, 刘伟, 尹小君, 等.基于无人机多光谱影像特征的最佳波段组合研究[J].农业机械学报, 2016(3): 242-248. http://www.cnki.com.cn/Article/CJFDTotal-NYJX201603034.htmZHAO Qingzhan, LIU Wei, YIN Xiaojun, et al. Research on optimal band combination based on multi spectral image characteristics of UAV[J]. Journal of Agricultural Machinery, 2016(3): 242-248. http://www.cnki.com.cn/Article/CJFDTotal-NYJX201603034.htm [9] REN Shaoqing, HE Kaiming, Girshick Ross, et al. Faster R-CNN: Towards Real-Time Object Detection with Region Proposal Networks[J]. IEEE Transactions on Pattern Analysis & Machine Intelligence, 2017, 39(6): 1137-1149. https://pubmed.ncbi.nlm.nih.gov/27295650/ [10] 郑国书.基于深度学习SSD模型的视频室内人数统计[J].工业控制计算机, 2017, 30(11): 48-50. http://www.cnki.com.cn/Article/CJFDTOTAL-GYKJ201711023.htmZHENG Guoshu. Statistics of indoor video population based on deep learning SSD model[J]. Industrial Control Computer, 2017, 30(11): 48-50. http://www.cnki.com.cn/Article/CJFDTOTAL-GYKJ201711023.htm [11] Oseph Redmon, Santosh Diwala, Ross Girshick, et al. You Only Look Once: Unified, Real-Time Object Detection[C]//IEEE Conference on Computer Vision and Pattern Recognition, 2016: 779-788. [12] 徐代, 岳章, 杨文霞, 等.基于改进的三向流Faster R-CNN篡改图像识别[J/OL].计算机应用, http://www.joca.cn/CN/10.11772/j.issn.1001-9081.2019081515.XU Dai, YUE Zhang, YANG Wenxia, et al. Tamper image recognition based on improved three-way Fasert R-CNN[J/OL]. Journal of Computer Applications, http://www.joca.cn/CN/10.11772/j.issn.1001-9081.2019081515. [13] 孙雄峰, 林浒, 王诗宇, 等.基于改进Faster RCNN的工业机器人分拣系统[J].计算机系统应用, 2019, 28(9): 258-263 https://kns.cnki.net/KCMS/detail/detail.aspx?dbcode=CJFD&filename=XTYY201909039SUN Xiongfeng, LIN Hu, WANG Shiyu, et al. Industrial robot sorting system based on improved fast RCNN[J]. Computer System Application, 2019, 28(9): 258-263. https://kns.cnki.net/KCMS/detail/detail.aspx?dbcode=CJFD&filename=XTYY201909039 [14] LONG J, Shelhamer E, Darrell T. Fully convolutional networks for semantic segmentation[C]//Computer vision and Pattern Recognition, 2015: 3431-3440. [15] Simonyan K, Zisserman A. Very Deep Convolutional Networks for Large-Scale Image Recognition[J]. Computer Science, 2014, 42(5): 324-328. https://arxiv.org/abs/1409.1556 [16] 张琦, 丁新涛, 王万军, 等.基于Faster R-cnn的交通目标检测方法[J].皖西学院学报, 2019, 35(5): 50-55. http://www.cnki.com.cn/Article/CJFDTotal-WXXB201905012.htmZHANG Qi, DING Xintao, WANG Wanjun, et al. Traffic target detection method based on fast R-cnn[J]. Journal of West Anhui University, 2019, 35(5): 50-55. http://www.cnki.com.cn/Article/CJFDTotal-WXXB201905012.htm [17] 陈泽, 叶学义, 钱丁炜, 等.基于改进的Faster R-CNN小尺度行人检测[J].计算机工程, 2020, 46(9): 226-232, 241.CHEN Ze, YE Xueyi, QIAN Dingwei, et al. Small scale pedestrian detection based on improved Faster R-CNN[J]. Computer Engineering, 2020, 46(9): 226-232, 241. [18] 侯志强, 刘晓义, 余旺盛, 等.基于双阈值-非极大值抑制的Faster R-CNN改进算法[J].光电工程, 2019, 46(12): 1-11. http://www.cnki.com.cn/Article/CJFDTotal-GDGC201912010.htmHOU Zhiqiang, LIU Xiaoyi, YU Wangsheng, et al. Improved Faster R-CNN algorithm based on double threshold-non-maximum suppression[J]. Optoelectronic Engineering, 2019, 46(12): 1-11. http://www.cnki.com.cn/Article/CJFDTotal-GDGC201912010.htm [19] HE K M, Gkioxari G, Dollár P, et al. Mask R-CNN[C]//Proceedings of IEEE International Conference on Computer Vision, 2017: 1026-1032. [20] Redmon J, Farhadi A. YOLOv3: An incremental improvement[J/OL]. arXiv: 1804. 02767, 2018.