基于单个摄像机的车辆检测与跟踪

【作者】 马伟；

【导师】 孙兴华；

【作者基本信息】 南京理工大学 ， 计算机应用技术， 2008， 硕士

【摘要】 本文的研究内容是基于单个摄像机的车辆检测与跟踪,主要是对本车前方的运动车辆进行检测和跟踪。车辆检测一般分为两个步骤:第一步,找到车辆的候选区域;第二步,对车辆候选区域进行确认。本文首先详细分析了前方车辆在图像中所具有的一些特征,最后融合了车辆底部的阴影、纹理、形状、和灰度对称特征进行车辆的检测。通过车辆底部的阴影找到感兴趣区域ROI(Region Of Interesting),然后运用熵初步过滤掉一些路面干扰区域,剩下的区域作为车辆检测的候选区域,完成车辆检测的第一步。由于模型匹配方法过分依赖模型,在模型匹配失败的时候无法完成车辆的检测,而且很难为车辆建立一个精确的模型,所以本文融合车辆的形状特征和灰度对称特征对车辆的候选区域进行确认,完成车辆检测的第二步。从实验结果看,本文所提出的多种特征融合的车辆检测方法能对前方不同形状、颜色、大小、距离的车辆进行检测,也能对不同姿态的车辆进行检测。而且本文的车辆检测算法受光照的影响小,能在一定程度上完成由于车辆的颠簸造成摄像机抖动情况下的车辆检测,也能完成复杂路况下的车辆检测,是一种非常有效的车辆检测方法。为了提高车辆检测算法的鲁棒性和减少算法的时间复杂度,车辆检测过后一般都需要对检测出的车辆进行跟踪。在对车辆进行跟踪的时候,本文首先分析了基于模型的运动车辆跟踪方法和基于Kalman滤波器的运动车辆跟踪方法。由于基于模型的运动车辆跟踪方法在进行车辆跟踪的时容易发生跟踪漂移,而且不能处理遮挡情况下的跟踪:基于Kalman滤波器的运动车辆跟踪方法一般都是假设车辆做匀速或者匀加速直线运动,对于实际车辆的运动规律不是很恰当,而且还存在Kalman滤波器参数初始化等问题,跟踪结果也不好。本文提出了基于模型和灰度对称融合的运动车辆跟踪方法。从实验结果看,基于模型和灰度对称融合的运动车辆跟踪方法能稳定的跟踪到前方远距离车辆,能适应光照剧烈改变情况下的车辆跟踪,能处理遮挡问题,是一种有效的车辆跟踪方法。更多还原

【Abstract】 The research content of this paper is vehicle detection and tracking based on a single camera. It is mainly to detect and track front vehicles. In general, vehicle detection is divided into two steps. The first step is to find vehicle candidate regions; the second step is to verify the vehicle candidate regions. In this paper, we firstly analyze the front vehicle’s features in detail, and then we fuse vehicle’s multiple features to detect a vehicle. The fused features include: shadow underneath the vehicle, texture, shape and gray-scale symmetry. According to find the shadow underneath a vehicle, we can get some ROI(Region Of Interesting), and then we use entropy to remove some disturbing regions which lack of entropy. The left regions are regarded as vehicle candidate regions. After above process, we finish the first step of vehicle detection. As model based method relies too much on model, it can not detect vehicles when model matches unsuccessfully. Another problem is that it is very difficult to establish a precise model for a vehicle. So we do not use model based method to detect a vehicle. We synchronously fuse vehicle’s shape and gray-scale symmetry features to verify vehicle candidate regions. After above process, we finish the second step of vehicle detection. So far, vehicle detection has finished after these two steps. Our experiment results show that vehicle detection based on fusion of multiple features proposed in this paper can detect front vehicles with different shape, color, size, distance and gesture. This method is affected little by light condition, and can detect vehicles to some extent when camera shocks. It also can detect vehicles under different complex road environments. The vehicle detection method proposed in this paper is an effective method.In order to improve the robustness of vehicle detection and reduce time consuming, it always needs to track the detected vehicles after vehicle detection step. In vehicle tracking step, we firstly use model based vehicle tracking method and kalman filter based vehicle tracking method to track the detected vehicles. The experiment results of these two methods are not good. Model based vehicle tracking method will cause tracking drift, and can not deal with occlusion; Kalman filter based vehicle tracking method generally assumes that the motion of vehicle is uniformly linear motion or uniformly accelerated linear motion. This assumption is not appropriate for the actual movement of vehicles, and there are still problems about initializing the parameters of kalman filter. For the above reasons, we propose a vehicle tracking method called vehicle tracking based on fusion of model and gray-scale symmetry. The experiment results show that the proposed method can stably track front vehicles with long distance. It can adapt the condition that light changes drastically, and can deal with occlusion. Vehicle tracking based on fusion of model and gray-scale symmetry is an effective vehicle tracking method.更多还原