Computer Engineering and Applications ›› 2021, Vol. 57 ›› Issue (3): 189-195.DOI: 10.3778/j.issn.1002-8331.1911-0163

Previous Articles     Next Articles

Improved R-SSD Panoramic Video Image Vehicle Detection Algorithm

WANG Dianwei, ZHAO Mengying, LIU Ying, SONG Haijun, XIE Yongjun   

  1. 1.School of Telecommunication and Information Engineering, Xi’an University of Posts and Telecommunications, Xi’an 710121, China
    2. Xi’an Institute of Optics and Precision Mechanics of Chinese Academy of Sciences, Xi’an 710119, China
  • Online:2021-02-01 Published:2021-01-29

改进的R-SSD全景视频图像车辆检测算法

王殿伟,赵梦影,刘颖,宋海军,谢永军   

  1. 1.西安邮电大学 通信与信息工程学院,西安 710121
    2.中国科学院 西安光学精密机械研究所,西安 710119

Abstract:

To address the issue that the performance of the Single Shot Multi-Box Detector(SSD) algorithm is suffering from low accuracy and high missed detection rate when detecting vehicle targets in panoramic video images, this paper constructs an improved SSD network architecture named R-SSD, and proposes a vehicle target detection algorithm in panoramic video image based on the proposed R-SSD. An RPN* network is added before the original SSD network, which can filter the prior box of negative samples, and roughly adjust the position and size of the prior box to provide good initial conditions for subsequent regression. A transmission conversion module is constructed between the original SSD and RPN* network to realize feature fusion between these two networks, and increase the low-level feature information, so as to improve the detection effect of the target. The comparative experiment results on panorama video image data set show that the proposed algorithm has a detection accuracy of 90.78%, which is significantly better than the SSD algorithm, and has higher detection rate and lower missing detection rate than the traditional SSD network in panoramic target vehicle detection.

Key words: panoramic vehicle object detection, Single Shot Multi-Box Detector(SSD) algorithm, feature fusion, transmission conversion module

摘要:

针对SSD算法在检测全景视频图像车辆目标时存在准确率低、漏检率高的问题,构建了一种改进的SSD网络,命名为R-SSD,并提出了一种基于R-SSD的全景视频图像中车辆目标检测算法。在原SSD网络之前增加了一个RPN*网络,目的在于过滤负样本先验框并粗略调整先验框的位置和大小,为后续回归提供好的初始条件。在原SSD和RPN*网络之间构建了传输转换模块,实现两个网络间的特征融合,并增加低层特征信息,从而提高目标的检测效果。在同时兼顾了RPN*网络和SSD*网络损失函数的基础上提出了新的损失函数,应用了二分类和多分类的方法,使回归操作更加精确。将采集的全景视频图像数据分为训练集和测试集,通过对比实验,表明提出的R-SSD算法检测精度可达90.78%,明显优于SSD算法,可较好地解决全景目标车辆检测中误检率较高、漏检率较高等问题。

关键词: 全景车辆检测, SSD算法, 特征融合, 传输转换模块