基于3D空间多部件模型的车辆检测方法研究

发布时间：2018-01-11 21:12

本文关键词：基于3D空间多部件模型的车辆检测方法研究　出处：《长安大学》2015年硕士论文　论文类型：学位论文

【摘要】：车辆检测是智能交通系统研究的一个重要领域。目前有多种车辆检测算法,如基于特征算子的方法、分类器法、3D建模法等,虽然取得了一定的效果,但是仍存在很多问题,如对环境变化的敏感、受目标遮挡和融合的影响等。本文考虑到三维空间中车辆特征,使用基于3D空间多部件模型的方法完成车辆检测。首先根据二维图像和三维空间的映射关系,提出在空间中虚拟一个逆投影面,然后利用空间到图像的逆透视投影恢复出该逆投影面上的信息,恢复的图称为逆投影图,逆投影图中目标的几何形状及尺寸特征都得到恢复。对车辆而言,根据夜晚和白天光照条件的不同,夜晚选取车头灯对作为车辆的固有部件,白天则选取车牌和尾灯对作为车辆的固有部件。接下来则使用不同的图像处理算法检测夜晚车头灯和白天车牌及尾灯,前者利用几何特征完成,后者利用颜色特征完成。部件空间关系概率模型符合高斯混合模型特征,从大量车辆样本中提取部件空间关系特征,用期望最大化算法完成模型的建立。最后,计算逆投影图中检测到的多个候选部件的概率似然度,大于预定阈值的则这些部件属于同一辆车。最终,多部件的检测结果就是车辆的检测结果。本文用多个交通场景视频对文中算法做了测试,采用的视频帧频为25帧/秒,大小为720*288。实验表明,该种方法适应性强,能应对光照条件差以及部分部件受遮挡的情况,还能避免相邻车辆的合并检测问题,并且对多种不同型号的车辆目标检测具有适用性。
[Abstract]:Vehicle detection is an important field in the research of intelligent transportation system. There are many vehicle detection algorithms, such as feature operator based method, classifier method and 3D modeling method, although some results have been achieved. However, there are still many problems, such as the sensitivity to environmental change, the influence of target occlusion and fusion, etc. In this paper, the characteristics of vehicles in three-dimensional space are considered. The method based on 3D space multi-component model is used to complete vehicle detection. Firstly, according to the mapping relationship between two-dimensional image and three-dimensional space, a virtual inverse projection plane is proposed in the space. Then the information on the inverse projection plane is restored by the inverse perspective projection from the space to the image. The restored graph is called the inverse projection map. The geometric shape and size characteristics of the object in the inverse projection map are restored. According to the different illumination conditions between night and day, the headlight pair is chosen as the inherent part of the vehicle at night. During the day, license plate and taillight pair are selected as the inherent parts of the vehicle. Then different image processing algorithms are used to detect the headlights at night and the license plates and taillights during the day. The former uses geometric features to complete the detection. The latter uses color features to complete. The probability model of component spatial relationship accords with Gao Si mixed model feature. The feature of component spatial relationship is extracted from a large number of vehicle samples and the model is built with expectation maximization algorithm. Finally. The probability likelihood degree of a plurality of candidate components detected in the reverse projection diagram is calculated, and if the probability likelihood degree is greater than the predetermined threshold value, the components belong to the same vehicle. The result of multi-component detection is the result of vehicle detection. This paper uses multiple traffic scene video to test the algorithm. The video frame frequency is 25 frames / second and the size is 720 / 288. the experiment shows that. This method has strong adaptability, can deal with the condition of poor illumination and some parts being occluded, and can avoid the problem of combined detection of adjacent vehicles, and it is applicable to the target detection of many different types of vehicles.
【学位授予单位】：长安大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：TP391.41;U495

【相似文献】