Road Scene Understanding
Perception plays a key role in any robotic application. In the case of intelligent vehicles, the perception task is referred to as road scene understanding. It involves using different sensors combined with automatic reasoning, in order to create a synthetic representation of the environment around the vehicle. The knowledge base accumulated by this task is then used either to issue warnings to the driver in the case of advanced driving assistance systems (ADAS) or to control vehicle actuators in the case of complete autonomous driving. A complete and precise description of the state of surrounding environment is the key factor that allows the reduction of the number of false and missed alarms and provides the basis for smooth automatic driving. Needless to say, the perception of an outdoor environment even if partially structured – is a challenging problem not only due to the intrinsic complexity of the driving environment itself, but also due to the impossibility of controlling many environmental parameters. Figure shows examples of day/night, sun/streetlight illumination, temperature, poor visibility, rain/snow, and different meteorological conditions, which in general are impossible to control and have to be faced by sensing devices.
The research community is addressing the issue of providing vehicles with robust and precise perception of the state of the environment from two different perspectives. One approach is to provide vehicles with ever-increasing sensing capabilities and processing power aimed at the provision of powerful onboard intelligent systems; Daimler-Chrysler is a world leader in this area. An alternative approach is to use road infrastructure as an active component capable of communicating with all vehicles and sharing information on road conditions in real time. Indeed these two perspectives can also be merged to provide a mixed solution to safely control a vehicle and in dynamic environments.
The task of road scene understanding may be addressed differently, depending on the availability of an intelligent infrastructure and on other players exhibiting cooperative behavior. The task of understanding the state of the environment can be simplified through the availability of information coming from other sources, thereby limiting the need to perform a robust and complete sensing on board each vehicle. Helpful information could come from the infrastructure itself (for example, road conditions and geometry, number of lanes, visibility, road signs, or even real-time information such as traffic-light status or traffic conditions) or other players (such as the presence of the vehicle with precise position, speed, and direction). The players may also carry real-time information gathered by and shared with other players.
Although research is currently focused on both intelligent vehicles and intelligent infrastructures, the first generation of production intelligent vehicles will have to rely primarily on their own sensing capabilities since the availability of information coming from other sources such as the infrastructure and other vehicles will take a while to be deployed in real-world situations. In fact, in order to be of practical use, intelligent roads must cover a large proportion of a country, and simultaneously cooperative intelligent vehicles must also be sufficiently widespread. It is important to note that the investment in intelligent infrastructures and intelligent vehicles comes from different sources: mainly from governmental institutions for the former, and vehicles owners for the latter.
The information that is owned by the infrastructure itself and that could be made available to the vehicles includes
（1）precise geometry of the lane/road
（2）road signs
（3）status of traffic lights
On the other hand, the infrastructure can also assess and deliver real-time data such as
（1）road conditions
（2）traffic conditions
Another important piece of information that needs to be gathered by intelligent vehicles is the presence of other road players, such as
（1）vehicles
（2）vulnerable road users (pedestrians, motorcycles, bicycles)
Although it could be assumed that sometime in the future all vehicles will be equipped with active systems that allow them to be safely avoided by other vehicles, it is quite improbable that pedestrians and bicycles will have similar equipments: their presence will need to be detected using onboard sensors only. The same consideration also applies to obstacles that may unexpectedly be found on the road, or to temporary situations such as road works: if a vehicle needs to cope with the unexpected, then it needs to have the capability to assess the situation in real time with its own sensors.
This is why onboard sensing is of paramount importance for future transportation systems; vehicle-to vehicle and vehicle-to-infrastructure communications may help and improve the sensing, but a complete sensor suite must also be installed on our future vehicles. The main challenges in road environment sensing are examined below.
Road/Lane Tracking
Many vehicles prototypes have been equipped with lane detection and tracking systems, starting from the very first implementation in the early 1980s. Indeed in this case computer vision plays a basic role; although generally the road can also be detected with laser scanners, the only generic technology able to detect lane geometry and lane markings with high precision is computer vision. Most lane tracking approaches have focused on detecting lane markings and exploiting structure in the environment, such as the parallelism of the left and right lane markings, the invariance of road width, or the widely used flat-road assumption. These assumptions were mainly used to overcome the problem of having a single camera (a choice driven by cost). Some systems use
剩余内容已隐藏，支付完成后下载完整资料
道路场景理解
感应在任何机器人应用中都起着重要的作用,在智能车辆的实例中,感应任务被称为道路场景理解。它涉及到使用不同的传感器并结合自动推理,来创造车辆周围环境的综合表示。以此任务积累的基础知识用来在先进辅助系统(ADAS)的情况下给驾驶员发出警告,或者在完全自动驾驶的情况下控制车辆的制动器。完整和准确的描述周围环境的关键因素是它允许减少虚报漏报的数量同时也为平滑的自动驾驶提供了基础。不用说的是,对室外环境的感应,即使是部分的结构化,也是一个很有挑战行的问题,这不仅是由于驾驶环境本身固有的复杂性,而且还由于控制许多环境参数的不可能性。图中展出了日/夜，太阳/路灯照明，温度，能见度差，雨/雪，和不同的气象条件的例子，这在一般是不可能控制，并且是传感装置必须要面对的。
研究团体正在从两个不同的角度来解决为车辆提供强大和精确的感知环境的状态。一种方法是为车辆提供不断增加的感测能力,目的提供强大的板载智能系统的处理能力,戴姆勒-克莱斯勒在这一领域是世界领先。另一种方法是使用道 路的基础设施作为能够与所有车辆进行通信和实时分享有关道路条件的信息的有效成分。的确这两种观点可以合并以提供一种混合的解决方案,可以在动态环境中安全的控制车辆。
道路场景理解任务可以用不同的方法解决,这取决于智能基础设施和其他工作者表现出的合作行为的可用性。理解环境状况的任务可以从其他来源的可用信息来简化。从而限制了需要执行一个健全和完整的感应每个车辆。有用的信息可能来自基础设施本身(例如,道路条件和几何、车道数量,可见性、道路标志、甚至实时信息,如红绿灯状态或交通条件)或其他用户(比如车辆的存在与精确位置,速度和方向)。用户也可能使用收集的实时信息和与其他用户共享。
虽然目前的研究集中在智能汽车和智能基础设施,生产的第一代智能车辆不得不主要依靠自己的感测能力,因为信息来自其他来源,如基础设施和其他车辆在现实世界中的部署情况,事实上,为了实际的应用,智能道路必须覆盖一个国家的大部分,同时智能汽车的合作也必须足够广泛。重要的是要注意,在智能基础设施和智能车辆的投资有着不同的来源,前者主要来自政府机构,后者来自车辆所有者。
属于基础设施本身的信息,可以为车辆提供包括:
bull;车道/路的精确几何
bull;路标
bull;交通信号灯的状态，另一方面,也可以评估的基础设施并提供实时数据等
bull;路况
bull;能见度
bull;交通状况
另一个重要的信息需要由智能车辆收集其他道路的信息，比如：
bull;车
bull;道路使用者（行人，摩托车，自行车）
尽管可以假定在未来某个时候，所有车辆将配备有源系统，使他们能够安全地避免其他车辆，很不可思议，行人和自行车会有类似的设备，它们的存在只需要机载传感器探测到。相同的这个方法也适用于在路上意外发现的障碍，或道路临时施工的情况。如果一个车辆需要应付意料之外的问题，那么它需要具有有能力实时评估情况的传感器。
这就是为什么机载遥感是未来非常重要的交通工具。车辆到车辆和车辆到基础设施通信可以帮助可提高传感，但一个完整的传感器套件还必须安装在我们的未来汽车上。在道路环境的主要挑战如下检查。
路/巷跟踪
许多车辆的原型已经配置了监测和跟踪系统，从20世纪80年代初的第一次实施。实际上在这种情况下计算机视觉起着基本的作用，虽然一般的道路也可用激光扫描仪探测。唯一的通用技术能够检测车道的几何形状和车道标记高精度的是计算机视觉检测。大多数车道跟踪方法主要集中在检测车道标记和利用结构环境，如并行的左和右车道标志，道路宽度的不变性，或广泛使用的扁平道路的假设。这些假设主要用于克服具有单个照相机（由成本驱动的选择）的问题。一些系统使用立体视觉检测车道标记，使工作没有这样的限制。在高速公路车道跟踪情况基本上是一个解决问题-与被部署在乘用车和商用车商业系统。一个典型的商业车道追踪例子如图所示。
然而,这样的系统不能保证车道探测系统有100%的可靠性,这些系统通常具有95-99%的可靠性。因此车道跟踪系统只正在车道偏离警告系统中使用，自主驾驶不能容忍错误，正在努力开发的新算法涵盖各种各样的驾驶环境，并推动着走向100%可靠的边界。
路标检测
另一种相当简单的计算机视觉的使用是路标检测和理解，路标指示牌是故意结构化，用来帮助司机。路标使用一组定义明确的形状，颜色和图案。路标放置在一致的高度跟位置有关。因此读路标是计算机视觉的一个可实现的任务。检测是使用形状或/和颜色检测方案的集合。经过检测和定位阶段，识别发生。通常这个任务是通过模式匹配技术来进行诸如图像互相关，神经网络，支持向量自组，因为可能集合的路标是有限的，明确界定。图中说明了基于高速标志检测超速预警系统概念。研究这方面的工作所面临的挑战在于检测的健全性和标志分类的可靠性。大多数汽车公司都在开发系统，如西门子。
红路灯检测
颜色和图案匹配，也是用于交通灯检测的关键技术。尽管交通灯的检测不是过于复杂，这个应用程序隐藏了另外一个方面，使车辆应用困难；除了正确的定位和信号的识别，必须注意检查信号的位置和方向，因为该信号通道可能不会被当前的车辆寻址到，尤其是市中心许多交通灯的十字路口在同一时间是可见的，车辆必须有能力选择正确和遵守的交通信号。一些实验已经实现有效的交通信号，能发射使用交通灯的状态的无线电频率。这涉及到额外的基础设施，在此阶段视觉似乎是唯一简单可行的解决方案。
能见度评估
其中一个关键挑战是雾的检测，国际照明委员会定义的气象能见度距离为超过一个适当大小的黑色物体被认为具有小于5%的对比度的距离限定。不同的技术测量该参数，因此检测大雾天气已经落实。虽然很多方法使用视觉，也有有效地替代品---通常用于固定位置，如机场和交通监测站---基于多次散射激光雷达的使用，利用视觉估计能见度的主要挑战是，行驶中的车辆在一个特定的距离一般不能依赖特定的参考点/对象/信号。
车辆检测
车辆的检测是使用多种传感器技术解决，从视觉到激光，雷达到声波。尽管在形状和颜色上不同，车辆公用相同的特点并配备了大尺寸的反光材料。一旦在路上/车道位置的粗略估计是可用的，那么车辆的位置是可以预见的，事实上，车辆可以通过许多不同的独立的传感器检测。图中所示是基于视觉的车辆检测系统。
然而,尽管解决这个问题看起来简单,每个传感器都有自己的应用程序域和它自己的挑战。想法是美好的,但在低能见度和不良照明场景(夜间或隧道)或在交通拥挤的条件下车辆可能咬合。视觉在红外域可以以高信置的检测到车辆目标，这是由于车辆轮胎和消音器通常显示出较高的温度，因此容易在图像中检测到。然而，停放的车辆，拖车，甚至刚刚开始移动的车辆比运行的车辆冷，因此会不太明显，激光雷达通常会很强劲，但在不良的条件下灵敏度会降低。雷达，虽然很便宜，可能会由于其他邻近反射物的存在会在横向测量时遭受偏压。最后声呐仅适用于短距离测量。该研究的最大挑战是实现多传感器的融合。一种常见的做法是将视觉与雷达融合。
行人检测
道路使用者（行人和自行车）的检测是智能汽车最困难的任务之一。行人的外观是很有挑战的：一个人的形状可以在几十毫秒内改变很大，但是颜色，纹理，或大小没有明显的变化，并且不能假设姿势，速度，或人体部分的可见性，如头部。机械学习的方法已经以跟高的可靠性成功的应用到这个问题。
高级驾驶员辅
剩余内容已隐藏，支付完成后下载完整资料
资料编号：[505494]，资料为PDF文档或Word文档，PDF文档可免费转换为Word
以上是毕业论文外文翻译，课题毕业论文、任务书、文献综述、开题报告、程序设计、图纸设计等资料可联系客服协助查找。
