一种对称帧差分约束的车辆自适应分割算法  

An Adaptive Vehicle Segmentation Algorithm Based on Symmetric Frame Differential Constraint

作　　者：吴宏涛 孟颖 雷凌昱 WU Hong-tao;MENG Ying;LEI Ling-yu(Shanxi Intelligent Transportation Research Institute Co.,Ltd.,Taiyuan,Shanxi 030032,China;Beijing GOTEC ITS Technology Co.,Ltd.,Beijing 100088,China)

机构地区：[1]山西省智慧交通研究院有限公司,山西太原030032 [2]北京中交国通智能交通系统技术有限公司,北京100088

出　　处：《公路交通科技》2024年第4期176-185,共10页Journal of Highway and Transportation Research and Development

基　　金：山西省重点研发计划项目(202102070301019);山西省基础研究计划项目(202103021223464);山西交通控股集团有限公司科技创新项目(23-JKKJ-20)。

摘　　要：为实现交通视频车辆目标的准确检测,提升交通事故的预防效果,提出一种对称帧差分约束下的交通视频车辆分割算法。首先,采用对称差分思想将相邻的三帧连续图像进行两次差分,提出基于自适应分布数约束的混合高斯模型背景选择更新方法,采用跳跃度和稳定度两个视频图像参数判断某一像素是否属于虚假目标点,通过背景选择更新机制实现前景目标的检测,利用自适应分布数的混合高斯模型算法提高高斯混合模型的计算效率,实现道路前景目标的准确分离。其次,对于结果图像,利用车辆阴影的特征和车辆边缘信息,提出一种基于车辆边缘修正的阴影检测与去除算法,增强复杂背景的车辆分割检测特征;然后,对于运动车辆目标通过自适应阈值进行分割,本研究在考虑类间方差的同时,考虑视频前景目标类的内聚性,将其作为阈值选取的一个标准,采用“最大方差比”自适应阈值选取算法,同时对于静止车辆目标通过背景更新筛选进行分割。最后,试验结果和对比分析表明,本研究提出算法可以将不同交通状态的车辆目标较完整、准确地分割出来,有效提高了交通场景下的不同状态车辆目标检测精度。In order to achieve accurate detection of vehicle targets in traffic videos and improve the prevention effect of traffic accidents,the vehicle segmentation algorithm for traffic video with the constraint of symmetrical frame difference is proposed.Firstly,the idea of symmetric difference is used to differentiate 3 adjacent consecutive images twice,and a background selection and update method with mixed Gaussian model based on adaptive distribution constraint is proposed to initially extract and update the background of high-definition video.Jump degree and stability in the image are used to determine whether a pixel belongs to a false target point.The background selection update mechanism is proposed to detect foreground targets,and the adaptive distribution number mixed Gaussian model algorithm is used to improve the computational efficiency of the Gaussian mixture model for achieving accurate separation of foreground targets.Secondly,for the resulting image,a shadow detection and removal algorithm based on vehicle edge correction is proposed,and vehicle segmentation detection features are enhanced under the complex background.Then,the moving vehicle target in the result image is segmented and detected by self-adaptive threshold.Besides inter class variance,the cohesion of video foreground target class is considered as a criterion for threshold selection.The adaptive threshold selection algorithm with maximum variance ratio is adopted,and the stopped vehicle is segmented by background updating and filtering.Finally,the experimental result shows that the proposed method can segment vehicle targets with different traffic scenes completely and accurately.In addition,the proposed method effectively improve the detection accuracy of vehicle targets with different situations in traffic.

关 键 词：智能交通 目标检测 背景更新 交通视频 自适应阈值 

分 类 号：U495[交通运输工程—交通运输规划与管理]