基于Kinect深度虚拟线圈的夜间车流量检测

• 作 者: 张汝峰 胡钊政 穆孟超
• 机 构: 河北工业大学电子与信息工程学院,武汉理工大学智能交通系统研究中心,
• 出 处: 交通信息与安全
• 发表日期: 2017年05期
• 关键词: 智能交通 夜间车流量检测 深度虚拟线圈 Kinect SVM 车型分类

针对夜间环境下基于摄像机的车辆检测方法存在精度低、稳定性差以及无法对车型进行有效识别等问题,提出一种基于Kinect深度虚拟线圈的夜间车辆检测与计数算法。首先对Kinect深度图像进行预处理,分别获得运动目标深度图(MDM)与空洞深度图(HDM)。然后在MDM与HDM上设置虚拟线圈,利用积分图像分别生成对应的一维运动信号,对其进行加权合成获得对车辆运动特征的表达,并在合成的运动信号范围内检测出车辆目标,并计算出车辆目标的几何特征,通过SVM对车型进行有效识别。实验结果表明,该算法对于单双车道的车辆计数正确率分别高达99.75%与99.25%,大小车型分类正确率可达99.80%,处理单张图片的平均时间仅为7ms。

Aimed at the problems of low precision, poor stability and no effective recognition of vehicle models for night-time camera-based vehicle detection methods, a night-time vehicle detection and counting algorithm based on Kinect depth virtual coils is proposed. Firstly, the Kinect depth image is preprocessed to obtain the MDM and HDM. Then a virtual coil is set up on the MDM and HDM, the corresponding one-dimensional motion signals are generated respectively by the integral images, the weighted synthesis is performed to obtain the expression of the vehicle motion features, and the vehicle targets are detected within the synthesized motion signals and calculated The geometric characteristics of the vehicle target, through the SVM to effectively identify the models. The experimental results show that the algorithm can accurately count 99.75% and 99.25% of single and double lane vehicles, respectively, and the correct classification rate of the vehicle models can reach 99.80%. The average time for processing single images is only 7ms.