-
摘要: 針對弱光照條件下交通標志易發生漏檢和定位不準的問題，本文提出了增強YOLOv3（You only look once）檢測算法，一種實時自適應圖像增強與優化YOLOv3網絡結合的交通標志檢測與識別方法。首先構建了大型復雜光照中國交通標志數據集；然后針對復雜的弱光照圖像提出自適應增強算法，通過調整圖像亮度和對比度強化交通標志與背景之間的差異；最后采用YOLOv3網絡框架檢測交通標志。為了降低先驗錨點框設置精度以及圖像中背景與前景比例嚴重失衡對檢測精度造成的影響，優化了先驗錨點框聚類算法和網絡的損失函數。對比實驗結果表明，在實時性大致相當的情況下，本文提出的增強YOLOv3檢測算法較標準YOLOv3算法對交通標志有更高的回歸精度和置信度，召回率和準確率分別提高0.96%和0.48%。Abstract: Traffic sign detection and recognition, which are important to ensure traffic safety, have been a research hotspot. In recent years, with the rapid development of automated driving technology, significant progress has been made in developing more accurate and efficient deep learning algorithms for traffic sign detection and recognition. However, these studies mainly focus on foreign traffic signs and do not consider the low-illumination conditions in practical application, which is a common scene. Therefore, many challenges still exist in the application of traffic sign detection and recognition in traffic scenes. To solve the problems of easy omission and inaccurate positioning for traffic sign detection and recognition under complex illumination conditions, the enhanced YOLOv3 (You only look once) detection algorithm, a traffic sign detection and recognition method combining real-time adaptive image enhancement and the YOLOv3 frame was proposed. First, a large and complex illumination traffic sign dataset for Chinese traffic was constructed; it included globally low illumination, locally low illumination, and sufficient illumination images. Then an adaptive enhancement algorithm was proposed for low-illumination images, which can enhance the difference between traffic signs and background by adjusting the brightness and contrast of the images. Finally, high-quality and discrimination images as input were transmitted to the YOLOv3 network framework, and traffic sign detection and recognition were performed. To reduce the influences of the prior anchor box setting accuracy and the imbalance between the background and foreground on the detection accuracy, the clustering algorithm for the prior anchor box and loss function for the network were optimized. The results of the comparison experiment with the LISA dataset and complex illumination traffic sign dataset for Chinese traffic show that the proposed enhanced YOLOv3 detection algorithm has higher regression accuracy and category confidence than the published YOLOv3 algorithm for traffic signs; the recall and precision are higher by 0.96% and 0.48%, respectively, which indicates the application potential of the proposed algorithm in actual traffic scenarios.
-
Key words:
- traffic sign detection /
- low illumination /
- adaptive image enhancement /
- YOLOv3 /
- deep learning
-
圖 1 不同天氣及光照條件的圖像樣本. （a）陰天; （b）雨雪天; （c）光照充足; （d）光照不足
Figure 1. Image samples under different weather and illumination conditions: (a) overcast; (b) rain and snow; (c) sufficient illumination; (d) insufficient illumination
圖 7 整體光照不足的圖像. （a）圖像處理前；（b）圖像處理后；（c）圖像處理前的像素概率直方圖；（d）圖像處理后的像素概率直方圖
Figure 7. Images with low overall illumination: (a) image before processing; (b) image after processing; (c) pixel probability histograms of image before processing; (d) pixel probability histograms of image after processing
圖 9 光照充足圖像。（a）圖像處理前；（b）圖像處理后；（c）圖像處理前像素概率直方圖；（d）圖像處理后像素概率直方圖
Figure 9. Images with sufficient illumination: (a) image before processing; (b) image after processing; (c) pixel probability histograms of image before processing; (d) pixel probability histograms of image after processing
圖 8 局部光照不足圖像。（a）圖像處理前；（b）圖像處理后；（c）圖像處理前像素概率直方圖；（d）圖像處理后像素概率直方圖
Figure 8. Images with low local illumination: (a) image before processing; (b) image after processing; (c) pixel probability histograms of image before processing; (d) pixel probability histograms of image after processing
圖 10 不同算法測試結果可視化對比. （a, b）標準YOLOv3；（c, d）改進YOLOv3
Figure 10. Visual comparison of different algorithm for test results: (a, b) standard YOLOv3; (c, d) improved YOLOv3
圖 11 不同算法測試結果可視化對比。（a, b）標準YOLOv3；（c, d）增強YOLOv3
Figure 11. Visual comparison of different algorithms for test results: (a, b) standard YOLOv3; (c, d) enhanced YOLOv3
表 1 圖像分類
Table 1. Image classification
Contrast category Intensity mean, λ Image category IL ≥ 0.5 Low contrast and high brightness < 0.5 Low contrast and low brightness IH ≥ 0.5 High contrast and high brightness < 0.5 High contrast and low brightness 
下載: 導出CSV
表 2 在LISA數據集上的測試結果（閾值=0.8，IOU=0.7）
Table 2. Test results on LISA dataset (threshold = 0.8, IOU = 0.7)
Algorithm Number of traffic signs Recall/% Accuracy/% Standard YOLOv3 1446 88.80 99.07 Improved YOLOv3 1446 91.36 98.44
下載: 導出CSV
中文字幕在线观看表 3 在弱光照交通標志數據集上的測試結果（閾值=0.8，IOU=0.7）
Table 3. Test results on weak illumination traffic signs dataset (threshold = 0.8, IOU = 0.7)
Algorithm Dark images with 2163 traffic signs Bright images with 2315 traffic signs All images with 4478 traffic signs Run time/ms Recall/% Accuracy/% Recall/% Accuracy/% Recall/% Accuracy/% Standard YOLOv3 96.30 98.91 98.49 99.30 97.43 99.11 33 Enhanced YOLOv3 98.06 99.44 98.70 99.74 98.39 99.59 36
下載: 導出CSV
-
參考文獻
[1] Zhang X Y, Gao H B, Zhao J H, et al. Overview of deep learning intelligent driving methods. <italic>J Tsinghua Univ Sci Technol</italic>, 2018, 58(4): 438張新鈺, 高洪波, 趙建輝, 等. 基于深度學習的自動駕駛技術綜述. 清華大學學報: 自然科學版, 2018, 58(4):438 [2] Salti S, Petrelli A, Tombari F, et al. Traffic sign detection via interest region extraction. <italic>Pattern Recognit</italic>, 2015, 48(4): 1039 doi: 10.1016/j.patcog.2014.05.017 [3] Eichner M L, Breckon T P. Integrated speed limit detection and recognition from real-time video // 2008 IEEE Intelligent Vehicles Symposium. Eindhoven, 2008: 626 [4] Lorsakul A, Suthakorn J. Traffic sign recognition using neural network on opencv: toward intelligent vehicle/driver assistance system. <italic>Intell Service Rob</italic>, 2007: 1 [5] Barnes N, Zelinsky A. Real-time radial symmetry for speed sign detection // IEEE Intelligent Vehicles Symposium. Parma, 2004: 566 [6] Loy G, Barnes N. Fast shape-based road sign detection for a driver assistance system // 2004 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). Sendai, 2004: 70 [7] Maldonado-Bascon S, Lafuente-Arroyo S, Gil-Jimenez P, et al. Road-sign detection and recognition based on support vector machines. <italic>IEEE Trans Intell Transportation Syst</italic>, 2007, 8(2): 264 doi: 10.1109/TITS.2007.895311 [8] Bahlmann C, Zhu Y, Ramesh V, et al. A system for traffic sign detection, tracking, and recognition using color, shape, and motion information // IEEE Proceedings. Intelligent Vehicles Symposium. Las Vegas, 2005: 255 [9] Girshick R, Donahue J, Darrell T, et al. Rich feature hierarchies for accurate object detection and semantic segmentation // 2014 IEEE Conference on Computer Vision and Pattern Recognition. Columbus, 2014: 580 [10] Uijlings J R R, van de Sande K E A, Gevers T, et al. Selective search for object recognition. <italic>Int J Comput Vision</italic>, 2013, 104(2): 154 doi: 10.1007/s11263-013-0620-5 [11] He K M, Zhang X Y, Ren S Q, et al. Spatial pyramid pooling in deep convolutional networks for visual recognition. <italic>IEEE Transactions Pattern Anal Mach Intell</italic>, 2015, 37(9): 1904 doi: 10.1109/TPAMI.2015.2389824 [12] Girshick R. Fast R-CNN // 2015 IEEE International Conference on Computer Vision (ICCV). Santiago, 2015: 1440 [13] Ren S Q, He K M, Girshick R, et al. Faster R-CNN: Towards real-time object detection with region proposal networks. <italic>IEEE Transactions Pattern Anal Mach Intell</italic>, 2017, 39(6): 1137 doi: 10.1109/TPAMI.2016.2577031 [14] Redmon J, Divvala S, Girshick R, et al. You only look once: Unified, real-time object detection // 2016 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Las Vegas, 2016: 779 [15] Liu W, Anguelov D, Erhan D, et al. SSD: Single shot multibox detector // European Conference on Computer Vision. Amsterdam, 2016: 21 [16] Ciresan D, Meier U, Masci J, et al. A committee of neural networks for traffic sign classification // The 2011 International Joint Conference on Neural Networks. San Jose, 2011: 1918 [17] Sermanet P, LeCun Y. Traffic sign recognition with multi-scale convolutional networks // The 2011 International Joint Conference on Neural Networks. San Jose, 2011: 2809 [18] Ciresan D, Meier U, Masci J, et al. Multi-column deep neural network for traffic sign classification. <italic>Neural Networks</italic>, 2012, 32: 333 doi: 10.1016/j.neunet.2012.02.023 [19] Jin J Q, Fu K, Zhang C S. Traffic sign recognition with hinge loss trained convolutional neural networks. <italic>IEEE Trans-actions Intell Transportation Syst</italic>, 2014, 15(5): 1991 doi: 10.1109/TITS.2014.2308281 [20] Mogelmose A, Trivedi M M, Moeslund T B. Vision-based traffic sign detection and analysis for intelligent driver assis-tance systems: perspectives and survey. <italic>IEEE Transactions Intell Transportation Syst</italic>, 2012, 13(4): 1484 doi: 10.1109/TITS.2012.2209421 [21] Stallkamp J, Schlipsing M, Salmen J, et al. Man vs. computer: Benchmarking machine learning algorithms for traffic sign recognition. <italic>Neural Networks</italic>, 2012, 32: 323 doi: 10.1016/j.neunet.2012.02.016 [22] Timofte R, Zimmermann K, Gool L V. Multi-view traffic sign detection, recognition, and 3D localization. <italic>Mach Vision Appl</italic>, 2014, 25(3): 633 doi: 10.1007/s00138-011-0391-3 [23] Yu C C. Research on Traffic Sign Detection and Recognition Algorithm[Dissertation]. Chengdu: Southwest Jiaotong University, 2017余超超. 交通標志檢測與識別算法研究[學位論文]. 成都: 西南交通大學, 2017 [24] Celik T, Tjahjadi T. Automatic image equalization and contrast enhancement using Gaussian mixture modeling. <italic>IEEE Trans Image Process</italic>, 2011, 21(1): 145 [25] Redmon J, Farhadi A. YOLOv3: an Incremental Improvement[J/OL]. arXiv preprint (2018-04-08)[2019-08-14]. http://arxiv.org/abs/1804.02767 [26] Redmon J, Farhadi A. YOLO9000: Better, faster, stronger // 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR). Honolulu, 2017: 6517 -
點擊查看大圖
計量
- 文章訪問數: 2122
- HTML全文瀏覽量: 1367
- PDF下載量: 100
- 被引次數: 0
