多维视差特征的复杂场景车辆检测关键技术研究

The stable and reliable vehicle detection is the primary task of traffic image analysis. This research obtains the disparity of image corresponding points and 3-D space information of vehicle detection through stereo vision technology to solve the scarcity of traffic parameters depth information. Therefore, this paper is focused on studying disparity image dimensionality reduction and cloud reconstruction method of complex wavelet pavement dense points, complex scene 3-D vehicle segmentation technology of structure dimension parallax projection and complex scene vehicle classification model of the depth dimension disparity characteristics. Thereby the rapid and robust 3-D vehicle detection and classification under complex road condition could be realized. Through fast Fourier Transform method on the basis of self-adaptive masking, this paper will solve the long matching time-consuming of binocular vision corresponding points and difficult application in real-time road scene. The vehicle occlusion and shadow interference caused by depth information deficiency could be solved by disparity characteristics 3-D vehicle segment method on the basis of relevant frequency domain. The study in this paper has significant theoretical significance and application value for complex scene vehicle detection method with multi-dimensional disparity characteristics.

稳定可靠的车辆检测过程是交通分析的首要任务。本课题利用立体视觉技术获得图像对应点的视差，获得检测车辆的三维空间信息，解决交通参数深度信息缺失的问题。为此，重点研究视差图像降维与复小波路面密集点云重构方法、结构维视差投影的复杂场景三维车辆分割技术和深度维视差特征的复杂场景车辆分类模型。从而实现复杂路况下快速、鲁棒的三维空间车辆检测及车型分类。课题通过基于自适应蒙版的快速傅里叶变换方法，解决双目视觉对应点匹配耗时巨大，很难在实时道路场景中应用的问题。通过基于频域相关的视差特征三维车辆分割方法，解决由于缺少深度信息，而造成的车辆遮挡及阴影干扰等问题。本课题研究在多维视差特征的复杂场景车辆检测方面具有重要的理论意义和应用价值。

车辆计数、车辆跟踪、车型分类、辅助驾驶、事故检测和道路行为判断都建立在准确的车辆检测基础上。本课题利用立体视觉技术获得图像对应点的视差，获得检测车辆的三维空间信息，解决交通参数深度信息缺失的问题。重点研究了视差图像路面密集点云重构方法、视差投影的复杂场景三维车辆分割技术和视差特征的复杂场景车辆分类模型。从而实现复杂路况下快速、鲁棒的三维空间车辆检测及车型分类。课题通过自适应匹配方法，解决双目视觉对应点匹配耗时巨大，很难在实时道路场景中应用的问题。通过基于频域相关的视差特征三维车辆分割方法，解决由于缺少深度信息，而造成的车辆遮挡及阴影干扰等问题。本课题研究在多维视差特征的复杂场景车辆检测方面具有重要的理论意义和应用价值。