Toward Perception-Driven Urban Environment Modeling for Automated Road Vehicles
Toward Perception-Driven Urban Environment Modeling for Automated Road Vehicles
复制标题
DOI：
10.1109/itsc.2015.124
复制
发表时间：
2015-09
期刊：
影响因子：
--
通讯作者：
J. Rieken;Richard Matthaei;M. Maurer
中科院分区：
文献类型：
--
作者：
J. Rieken;Richard Matthaei;M. Maurer
自动驾驶是当今广泛讨论的话题。令人印象深刻的演示展示了车辆自动化的潜力。然而，在自动驾驶的背景下，许多项目使用先验数据，以弥补在感知和了解车辆环境方面的不足。此外，在功能安全性和冗余性方面，目前尚不清楚这种基于本地化和地图的方法是否真的是开创性的。这就是为什么我们关注车载感知静止的城市环境的原因。虽然目标跟踪是一种常用的方法，但基于网格和基于对象的环境感知表示的结合仍然是一个研究课题。对车道和可行驶区域的充分感知是城市环境中一个悬而未决的问题。几个感知模块必须协作，以适当地表示车辆的周围环境。在这篇文章中，我们介绍了StadtPilot项目对感知驱动的城市环境建模的最新贡献。提出了一种基于栅格表示不同环境特征的车道检测方法。我们的方法能够检测多车道结构，并且能够处理复杂的车道结构，这是典型的城市道路。提取的特征由跟踪模块稳定。此外，我们结合了自由空间表示，其数据不是隐含地从检测到的目标获得的，而是基于显式的地面表示。动态分类模块的扩展关注于其他道路使用者的起停行为，以增强轨迹列表(移动对象)和网格(静止环境)的完备性。所提出的算法在标准PC上实时运行，并用真实的传感器数据进行了评估。
Automated driving is a widely discussed topic nowadays. Impressive demonstrations have shown the potentials of vehicle automation. However, many projects in the context of automated driving use a priori data in order to compensate insufficiencies in perceiving and understanding the vehicle's environment. Additionally, in terms of functional safety and redundancy, it is not yet known whether such localization-and map-based approaches are really path breaking. This is the reason why we focus on on-board perception also of the stationary urban environment. While object tracking is a commonly used approach, the combination of grid-based and object-based representations for environment perception is still a research topic. The sufficient perception of lanes and drivable areas is an unsolved issue in urban environment. Several perception modules have to collaborate for a suitable representation of the vehicles' surroundings. In this paper, we present the latest contributions of the project Stadtpilot to a perception-driven modeling of urban environments. We propose a lane detection approach which is based on a grid-based representation of different environmental features. Our approach is able to detect multi-lane structures and it is capable to deal with complex lane structures which are typical of urban roads. The extracted features are stabilized by a tracking module. Additionally, we incorporate a free-space representation which data is not derived implicitly from detected targets, but based on an explicit ground representation. Extensions of our dynamic classification module focus on the start/stop behavior of other road users in order to enhance the completeness of track list (mobile objects) and grid (stationary environment). The presented algorithms run in real-time on a standard PC and are evaluated with real sensor data.