计算机工程与应用 ›› 2020, Vol. 56 ›› Issue (22): 173-184.DOI: 10.3778/j.issn.1002-8331.2004-0379

• 模式识别与人工智能 • 上一篇    下一篇

改进YOLOv3的车辆实时检测与信息识别技术

顾恭,徐旭东   

  1. 北京工业大学 信息学部 计算机学院,北京 100124
  • 出版日期:2020-11-15 发布日期:2020-11-13

Real-Time Vehicle Detection and Information Recognition Technology Based on YOLOv3 Improved Algorithm

GU Gong, XU Xudong   

  1. School of Computer Science, Beijing University of Technology, Beijing 100124, China
  • Online:2020-11-15 Published:2020-11-13

摘要:

在复杂无约束自然场景下对车辆实时检测和相关信息的提取识别一直是计算机视觉领域内重要的研究内容之一。该领域问题的突破不但可以为汽车自动驾驶技术的实现和完善带来实际效果的提升,并且在停车场的自动停车调度算法和实时泊车监控系统的改进上有着重要的现实意义。针对当前实时车辆信息检测中存在的车辆检测区域不完整、精度不高以及无法对场景中较远车辆进行准确定位等相关问题,提出了一种Vehicle-YOLO的实时车辆检测分类模型。该模型在最新的YOLOv3算法基础上,通过更改图像输入参数,增强深度残差网络的特征提取能力,采用5个不同尺寸的特征图依次对潜在车辆的边界框提取等方式来提升车辆实时信息检测的精度和普适性,并通过KITTI、VOC等数据集进行性能验证和分析。实验结果表明,Vehicle-YOLO模型在KITTI数据集上达到了96%的均值平均精度,传输速度约为40 f/s,在精度提升的情况下仍能保持良好的实时检测速率。此外,Vehicle-YOLO检测模型在VOC等其余数据集上的实验结果也展现了不同程度的精度提升,故该模型在常见物体的定位检测中有较好的普适性,相较于传统的物体检测算法模型有更好的表现。

关键词: 车辆实时检测, YOLOv3, 目标定位, 卷积神经网络, 深度残差网络, 特征图

Abstract:

Real-time detection and extraction of vehicle information in complex natural scenes have always been one of the important contents in the field of computer vision. The breakthrough of problems in this field can not only bring effective promotion to the realization of auto driving technology, but also have important practical significance in the improvement of real-time parking monitoring system and automatic parking schedule algorithm in parking lot. In order to solve the problems of incomplete vehicle information detection area, low accuracy and inability to locate distant vehicles in real-time vehicle information detection, a new real-time vehicle detection and classification model—Vehicle-YOLO is proposed. Based on the latest YOLOv3 algorithm model, the model improves the accuracy and universality of vehicle real-time information detection by changing image parameters, enhancing features extraction ability of deep residual network, and using 5 different scale convolutional feature maps to extract potential bounding boxes of vehicles. Then, the odel performance is tested and analyzed with datasets, such as KITTI and VOC. Experimental results indicate that the Vehicle-YOLO model achieves 96% mean average accuracy and transmission speed about 40 f/s on KITTI dataset, in other words, the model can still maintain a better real-time detection speed when detection accuracy is improved. In addition, the experimental results of Vehicle-YOLO on other datasets such as VOC also show different degrees improvement of accuracy, which demonstrate that the proposed model has better universality and performance in common object positioning detection than traditional algorithms model of object detection.

Key words: real-time vehicle detection, YOLOv3, target detection, convolutional neural network, deep residual network, feature map