Night vehicle detection method based on improving Mask RCNN
-
摘要: 传统的夜间车辆检测基于车灯特征的提取和识别，这类方法容易发生误判、检测精度和检测实时性不高。针对上述问题，本文研究了基于改进Mask RCNN（mask RCNN-night vehicle detection，Mask RCNN-NVD）的夜间车辆检测算法。将残差网络（residual network，ResNet）结构中的普通卷积修改为数量为16组的分组卷积，通过16组1×1卷积实现通道数叠加，将网络参数降至普通卷积的1/16，提升检测速度，并实现与普通卷积相同的效果；将通道注意力机制模块（squeeze-and-excitation，SE）嵌入ResNet结构中，通过2个全连接层构建瓶颈结构，将归一化权重加权到各通道特征，增强网络表征能力；在特征金字塔网络（feature pyramid networks，FPN）后加入自底向上结构，将底层特征强定位信息传递到高层语义特征中；加入自适应池化层，根据区域候选网络（region proposal network，RPN）产生的候选区域分配至不同尺度特征图中，并在底层特征与各阶段最高层特征之间加入跳跃连接结构，实现缩减模型参数的同时保留模型的全局表征能力。通过对开源数据集Microsoft common objects in context（MS COCO）、Berkeley deep drive 100K（BDD100K）的夜间行车图像进行数据增强，构建用于评估检测性能的测试集2 000张。实验结果表明：算法在测试集上的平均精度（mean Average Precsion，mAP）值高达92.62，每秒图像处理帧数（Frames Per Second，FPS）值高达30帧。相比于原始Mask RCNN算法分别在mAP值上提高1.68，FPS值提高4帧，验证提出的方法可以有效提升夜间车辆检测的准确性和实时性。
-
关键词:
- ADAS驾驶辅助系统 /
- 夜间车辆检测 /
- Mask RCNN-NVD检测算法 /
- 深度学习
Abstract: The traditional nighttime vehicle detection method is generally based on the extraction and identification of headlights, which is prone to misjudgment, detection accuracy, and real-time lack of deficiency. To address the above issue, a night vehicle detection algorithm based on improved Mask RCNN night vehicle detection (Mask RCNN) is studied. The normal convolution in the residual network (ResNet) structure is modified to a grouped convolution of 16 groups, and channel number superposition is achieved by 16 groups of 1×1 convolution Network parameters are reduced to 1/16 of a normal convolution. The detection speed is improved and achieves the same effect as normal convolution. The channel attention mechanism module (squeeze-and-excitation, SE) is embedded in the ResNet structure, two fully connected layers are used to build the bottleneck structure, normalized weights are weighted to each channel feature, The representational power of the network is enhanced; Bottom-up structures are added behind feature pyramid networks (FPNs), the strong localization information of the underlying features is passed to the high-level semantic features; Adaptive pooling layer is added. The region proposal network (RPN) generates candidate regions, which are subsequently assigned to feature maps at different scales by region. The jump connection structure is added between the bottom feature and the top feature at each stage, The model parameters are reduced, while retaining the global representational power of the model. The open-source dataset Microsoft common objects in context (MS COCO), and Berkeley Deep Drive 100K (BDD100K) contain some nighttime driving images, these images are image enhanced. A test set is constructed to evaluate the performance of the detection, it contains 2 000 images. The Mask RCNN-NVD algorithm is tested on the test set, the mean Average Precision of Mask RCNN-NVD is 92.62, and the Frames Per Second (FPS) of Mask RCNN-NVD is 30 frames. Compared with the original Mask RCNN algorithm, the mapped value is improved by 1.68 and the FPS value is improved by 4 frames. The proposed method is validated, and nighttime vehicle detection is improved in both accuracy and real time. -
图 1 ResNet残差块普通卷积与ResNet-SE残差块分组卷积
Figure 1. ResNet residual block normal convolution and ResNet-SE residual block group convolution
图 11 存在对向来车和同向有车的场景下的检测效果
Figure 11. Detection effect in the scene of opposite direction vehicles and vehicles in the same direction
表 1 实验参数
Table 1. Experimental parameters
参数 取值 迭代轮数 300 输入数据批次 4 输入图像尺寸 800×600 初始学习率 0.001 动量参数 0.9 
下载: 导出CSV
表 2 Mask RCNN与Mask RCNN-NVD混淆矩值对比
Table 2. Comparison of Mask RCNN and Mask RCNN-NVD confusion moment values
算法 Tp Fp Tn Fn Mask RCNN 2 450 203 215 259 Mask CNN-NVD 2 528 151 186 207
下载: 导出CSV
表 3 不同算法实验结果对比
Table 3. Comparison of experimental results of different algorithms
算法 精确率/% 召回率/% mAP FPS/（帧/s） YOLOv3 83.65 81.19 81.92 36 SSD300 83.61 77.38 81.10 38 Faster RCNN 90.62 86.67 88.49 25 Mask RCNN 92.36 90.41 90.94 28 Mask RCNN-NVD 94.36 92.42 92.62 32
下载: 导出CSV
-
[1] 胡忆平. 基于深度学习的夜间车辆检测算法研究[D]. 长沙: 湖南大学, 2020.HU Y P. Research on nighttime vehicle detection algorithm based on deep learning [D]. Changsha: Hunan University, 2020. (in Chinese) [2] 李慧, 白鹏飞, 李世晓, 等. 基于车灯特征的夜间车辆检测方法[J]. 电子测量技术, 2020, 43(14): 89-95. doi: 10.19651/j.cnki.emt.2004290LI H, BAI P F, LI S X, et al. Nighttime vehicle detection method based on headlight features[J]. Electronic Measurement Technology, 2020, 43(14): 89-95. (in Chinese) doi: 10.19651/j.cnki.emt.2004290 [3] WU B F, CHEN Y L, CHIU C C. A discriminant analysis based recursive automatic thresholding approach for image segmentation[J]. IEICE Transactions on Information and Systems, 2005, 88(7): 1716-1723. [4] PARVIN S, ROZARIO L J, ISLAM M E. Vision-based on-road nighttime vehicle detection and tracking using taillight and headlight features[J]. Journal of Computer and Communications, 2021, 9(3): 29-36. doi: 10.4236/jcc.2021.93003 [5] KAVYA T S, TSOGTBAATAR E, JANG Y M, et al. Night-time vehicle detection based on brake/tail light color[C]. 2018 International SoC Design Conference(ISOCC), Daegu, Korea: IEEE, 2018. [6] AHMAD A, JACK A P, ARLAN A G. Detection and classification of vehicles for traffic video analytics[J]. Procedia Computer Science, 2018(144): 259-268. [7] MIAO Y, LIU F, HOU T, et al. A nighttime vehicle detection method based on yolo v3[C]. 2020 Chinese Automation Congress(CAC), Shanghai, China: IEEE, 2020. [8] HUANG S, HE Y, CHEN X. M-YOLO: A nighttime vehicle detection methodcombining mobilenet v2 and yolo v3[J]. Journal of Physics: Conference Series, 2021, 1883(1): 012094. doi: 10.1088/1742-6596/1883/1/012094 [9] HE K, GKIOXARI G, DOLLAR P, et al. Mask r-cnn[C]. The IEEE International Conference On Computer Vision, Venice, Italy: ICCV, 2017. [10] LOU L, ZHANG Q C, LIU M, et al. Detecting and counting the moving vehicles using mask R-CNN[C]. IEEE 8th Data Driven Control and Learning Systems Conference, Dali, China: IEEE, 2019. [11] 杨鹏强, 张艳伟, 胡钊政. 基于改进RepVGG网络的车道线检测算法[J]. 交通信息与安全, 2022, 40(2): 73-81. doi: 10.3963/j.jssn.1674-4861.2022.02.009YANG P Q, ZHANG Y W, HU Z Z. Lane-line detection algorithm based on improved RepVGG network[J]. Journal of Transport Information and Safety, 2022, 40(2): 73-81. (in Chinese) doi: 10.3963/j.jssn.1674-4861.2022.02.009 [12] PANT G, YADAV D P, GAUR A. ResneXt convolution neural network topology-based deep learning model for identification and classification of pediastrum[J]. Algal Research, 2020, 48: 101932. doi: 10.1016/j.algal.2020.101932 [13] WANG K, LIEW J H, ZOU Y, et al. Panett: few-shot image semantic segmentation with prototype alignment[C]. 2019 IEEE/CVF International Conference on Computer Vision, Seoul, South Korea: ICCV, 2019. [14] LIU W, ANGUELOV D, ERHAN D, et al. Ssd: single shot multibox detector[C]. Computer Vision-ECCV 2016: 14th European Conference, Amsterdam, Netherlands: ECCV, 2016. [15] SUN X, WU P, HOI S C H. Face detection using deep learning: An improved faster RCNN approach[J]. Neurocomputing, 2018(299): 42-50. [16] 刘勃, 周荷琴, 魏铭旭. 基于颜色和运动信息的夜间车辆检测方法[J]. 中国图象图形学报: A辑, 2005, 10(2): 187-191. https://www.cnki.com.cn/Article/CJFDTOTAL-ZGTB200502008.htmLIU B, ZHOU H Q, WEI M X. Night vehicle detection method based on color and motion information[J]. Chinese Journal of image and graphics: Series A, 2005, 10(2): 187-191(in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-ZGTB200502008.htm [17] 祁秋红, 陈启兴. 基于尾灯跟踪的夜间车辆检测[J]. 通信技术, 2012, 45(10): 58-60. https://www.cnki.com.cn/Article/CJFDTOTAL-TXJS201210020.htmQI Q H, CHEN Q X. Night vehicle detection based on tail light tracking[J]. Communication technology, 2012, 45(10): 58-60. (in Chinese) https://www.cnki.com.cn/Article/CJFDTOTAL-TXJS201210020.htm [18] VARGAS M, MILLA J M, TORAL S L, et al. An enhanced background estimation algorithm for vehicle detection in urban traffic scenes[J]. IEEE Transactions on Vehicular Technology, 2010, 59(8): 3694-3709. [19] LIU Y, QIU T, WANG J, et al. A nighttime vehicle detection method with attentive GAN for accurate classification and regression[J]. Entropy, 2021, 23(11): 1490. [20] LEUNG H K, CHEN X Z, YU C W, et al. A deep-learning-based vehicle detection approach for insufficient and nighttime illumination conditions [J]. Applied Sciences, 2019, 9(22): 4769. [21] 周君, 包旭, 高焱, 等. 基于模式识别与ST-MRF相结合的车辆检测方法[J]. 交通信息与安全, 2021, 39(2): 95-100. doi: 10.3963/j.jssn.1674-4861.2021.02.012ZHOU J, BAO X, GAO Y, et al. A vehicle detection method based on pattern recognition combined with ST-MRF [J]. Journal of Transport Information and Safety, 2021, 39(2): 95-100. (in Chinese) doi: 10.3963/j.jssn.1674-4861.2021.02.012 -
点击查看大图
图(11) / 表(3)
计量
- 文章访问数: 801
- HTML全文浏览量: 406
- PDF下载量: 36
- 被引次数: 0