【作者】 李仪；

【导师】 蔡自兴；

【作者基本信息】 中南大学 ， 计算机应用技术， 2013， 博士

【摘要】 无人驾驶智能车辆在城市环境中行驶时需要对图像数据中的道路、前方车辆、前方车辆的刹车灯和转向灯以及交通信号灯进行检测与识别。一方面无人驾驶智能车辆自身空间的限制导致其计算能力有限,另一方面图像数据由于摄像头的发展导致其数据量日益增大,且由于图像处理算法的复杂度逐渐提高其需求更多的计算资源,两者之间难以取舍；解决此矛盾的最好方法即建立混合异构并行计算模型来充分利用其计算资源。总而言之,本文做了如下工作：1.针对图像分割中的阈值难以确定,本文设计了多类类间最大方差法中的分类数确定算法,以此来获取多个阈值及分割图像；设计了包含道路区域的二值图像选择算法和基于二次多项式的分段函数拟合方法,基于该两个方法来估计道路区域的边界；为了定位前方车辆在图像中的位置,可以通过分析车底阴影的形状设计约束条件来选择车底阴影区域。通过实验表明该方法有较好的效果。2.利用车辆尾灯的镜像特征设计车尾灯对匹配算法,并通过其与车底阴影之间的位置关系来检测车尾灯对以及辅助定位车辆位置。为了提高检测率,设计了基于对象相对位置关系的多目标Kalman滤波跟踪算法。为了获得前方车辆的减速信息,设计了高位刹车灯检测,并通过刹车灯在“亮”状态下的颜色分布设计了刹车灯点亮状态识别算法,结合该两个算法的结果来识别前方刹车灯的状态。转向灯有左右之分并且以断续闪烁过程来提供给后方驾驶员转向信息。本文设计了当车尾灯对检测失效时的左右转向灯的判别方法,并设计了转向灯断续闪烁状态累计识别算法来识别转向灯是否处于“灭”—“亮”—“灭”这种状态转换的过程中。实验结果表明了这些方法的有效性。3.根据交通信号灯的光源部分具备“亮”特点,而其背板部分具备“暗”特点,设计了基于Top-hat算子的“亮”区域提取算法和基于多阈值分割的“暗”区域提取算法,通过两个区域的位置匹配来定位交通信号灯。对圆形光源区域设计了结合圆形度检测和改进Hough椭圆检测的方法来判断其是否为圆形；对箭头形光源区域,建立了标准箭头形状的面积投影函数,并对区域样本做投影转换及归一化,通过样本面积投影函数与标准面积函数的方差来确定方向。通过实验分析验证了这些算法的有效性。4.介绍了混合异构车载并行计算模型,对多台车载联网计算机设计了其拓扑结构以及基于在线动态分析的任务分配与作业调度算法,提出了GPU+CPU负载均衡优化策略减少了TBB与CUDA之间的数据交换。针对整体检测与识别系统进行总体任务划分以实现层级化管理,并对每一个任务进行了算法级的并行设计,以此实现了粗粒度以及细粒度上的混合异构并行设计。通过实验表明,一方面解决了图像处理的实时性问题,另一方面解决了资源有效利用的问题。更多还原

【Abstract】 Driving in city enviroments, the unmanned intelligent vechile needs to detect and recognize the objects in the image data, such as road, front vehicle, brake light and turn light of front vehicle, traffic light and so on. On the one hand, because of space limit, the unmanned intelligent vehicle’s computing combability is limited; on the other hand, owing to the development of camera device, image data is bigger than past, and with the increasingly complex, image processing algotithm needs more computing resource. It is hard to decide between them. But, the best way to resolve this problem is to construct the hybrid heterogeneous parallel computing model to utilize the computing resource sufficiently. All in all, in this paper, some works have been done as below.1. Due to the threshold of segmentation is hard to get, a number of classes selection algorithm in multi-class maximum variance menthod is designed to get multi-thresholds for segmentation. Binarization image including road region selection algorithm and on the basis of quadratic polynomial piecewise function fitness method are designed to estimate the edge of road. According to analysis of shape of the shadow underneath vehicles, some constraint conditions are designed for detecting the positions of front vehicles in image. Experiment results shows that these methods’ availability.2. By the mirror feature of real lamps, pairing real lamps detection algorithm is designed, and the shadow of front vehicles are cooperated with it for locating the vehicles. According to the relative relation among positions of regions, a multi-objects Kalman filter tracking algorithm is designed for increasing the detection rate of pairing rear lamps. In order to get the deceleration message of front vehicle, two algoritm are combined to detect and recognize. One is the high location brake light detection algorithm; the other is brightness state of brake light recognition algorithm according to the distribution of the color when the lamp is lit. When the left or right turn lamps are flashing, them send different messages to the rear drivers, respectively. If the pairing rear lamps detection is invalid, an algorithm named turn lamps left-right diviced method is designed. To recognize if the turn lamps being in the process of "extinguished"-"lit"-"extinguished" states conversion, a turn lamp flashing state accumulated recognition algorithm is designed. Experiments results show that these algorithms are effective.3. According to the light source of traffic light which has the feature of "bright", and back board of it which has the feature of "dark", two algorithms are designed to abstract the regions which are matched each others for locating the traffic lights. The one is "bright" regions abstraction algorithm based on Top-hat operator, and the other is "dark" regions abstraction algorithm based on multi-thresholds segmentation.To recognize if the regions are circular light source of circle traffic light, circurity method and improved Hough circle detection are designed. The standard arrow shape area projected functions are established. Then, the sample of region is projected and normalized. The variance between the reality value and estimated value is the key criterion to recognize if the regions are arrow light source. Experiments results show that these algorithms are effective.4. A hybrid heterogeneous on-board parallel computing model is introduced. The topological structure of on-board networked computers is constructed, and the task distribution and Job scheduling algorithm on the basis of on-line dynamic analysis is designed. The GPU+CPU load balancing optimization strategy is presented to reduce the data exchange. The whole detection and recognition system is divided by tasks for stratification management, and every tasks is parallel designed at algorithm level. The processes above accomplish hybrid heterogeneous parallel design at both coarse-grained and fine-grained.更多还原