本文为荷兰代尔夫特理工大学（作者：C.D. Berger）的硕士论文，共84页。

在过去的几年中，许多SAE 2级驾驶自动化系统已经出现在商用车中。沃尔沃的驾驶员辅助系统就是一个例子，当动态驾驶任务的某些部分被取消时，它为驾驶员提供了一个更舒适的旅程，但它仍然要求驾驶员注意，以便监督系统并在必要时恢复车辆运行。世界各地的汽车制造商和高科技公司都在致力于开发一种3级或更高级别的自动驾驶系统（ADS），驾驶员不需要监督，但迄今还未能向公众推出任何一种。在空旷的高速公路上驾驶车辆相对容易，但道路越来越拥挤，因此密集的交通驾驶也需要考虑在内。其中一个问题是，无监督驾驶不仅要求系统本身能够正确操作车辆，而且要求它能够处理其他驾驶员的错误。为了考虑这些以及其他不可预测的事件，当前的驾驶自动化系统倾向于采用相对于周围交通的较大纵向裕度，以利用其制动能力避免碰撞。在交通拥挤的情况下，这种保守的行为禁止改变车道。

本论文与Volvo Cars/Zenuity合作开发了一种车道变换路径规划算法，该算法能够在不影响避免可能发生碰撞的情况下，通过驶入间隙来规划密集交通中的主动车道变换。沃尔沃早期的研究表明，不仅考虑车辆能够刹车，还能够进行机动规避。如果可以使用规避机动，则与“仅制动”情况相比，必要的安全裕度会大大降低。在这项工作中，随着规避机动的建模和约束公式的发展，如果能够满足条件，那么能够保证设计规避机动的可用性。此约束转化为状态相关的安全区，ADS应始终保持车辆的安全距离。

模型预测控制通过将其视为优化的一个约束条件来实现这样一个安全区，从而保证在预测视界的任何一点上都可以进行规避机动。这就否定了单独规划所有可能规避机动的必要性，因此选择它作为路径规划方法。在模型预测控制框架中，对多部周围车辆实施多车道的安全区约束，以及在预测范围内超越其他车辆的要求，需要开发一种新的三步算法。该算法连续求解三个最优控制问题，在不突破安全区的情况下，规划出一条最优路径，使ego车辆在目标间隙内定位。为了能够实时实现该算法，选择了一个计算方便的车辆模型，并采用最先进的结构（利用FORCES-Pro生成的高效C代码）求解OCP。在这种设置下，以100[ms]采样的10[s]预测视界计算时间约为30[ms]。仿真结果表明，该算法能够找到最大侧向切入点，在安全区边界上进行操作，并能在车辆被后方切断或前方车辆突然停车时避免碰撞。结果表明，在不影响安全的情况下，ADS可以在密集的交通流中对车辆进行操作。通过使用已开发的安全区和新颖的解决方案框架，可以减少必要的裕度，从而提高车道变换机动的可用性。此外，采用所需的控制策略为实现实时适用性铺平了道路，这样的研究有助于下一步开发一个实际可用的自动驾驶系统。

In the past years a number of SAE level 2driving automation systems have come available in commercial vehicles. Anexample is Volvo’s Pilot Assist, which provides a more comfortable journey forthe driver as certain parts of the dynamic driving task are taken away, but itstill requires the driver to be attentive in order to supervise the system andresume vehicle operation if necessary. Car manufacturers and high-techcompanies all around the world are working on the development of an AutomatedDriving System (ADS) of level 3 or higher where the driver does not need tosupervise, but have so far not been able to introduce any to the public.Operating a vehicle on an empty highway is relatively easy, but roads aregetting more and more congested so dense traffic driving needs to be accountedfor as well. One of the issues is that unsupervised driving does not onlyrequire that a system itself operates the vehicle correctly, but requires thatit is able to cope with mistakes of other drivers. In order to take these intoaccount as well as other unpredictable events, current driving automationsystems tend to adopt large longitudinal margins with respect to surroundingtraffic to use their ability to brake to avoid a collision. In dense trafficsituations this conservative behaviour prohibits the ADS from making lanechanges. This thesis was setup in cooperation with Volvo Cars / Zenuity todevelop a lane change path planning algorithm that can plan active lane changesin dense traffic, by pushing into a gap without compromising the ability toavoid a possible collision. Earlier work at Volvo suggested considering theability of a vehicle to not only brake, but also to make an evasive maneuver.If an evasion maneuver is available the necessary safety margins are reducedconsiderably compared to a ‘braking-only’ scenario. In this work, the evasivemaneuver was modeled and a constraint formulation was developed that if met,guarantees the availability of the designed evasion maneuver. This constrainttranslates to a state-dependent safety zone which the ADS should keep thevehicle clear off at all times. Model Predictive Control allows theimplementation of such a safety zone by considering it as a constraint on theoptimisation, guaranteeing that an evasion maneuver is available along anypoint on the prediction horizon. This negates the necessity to plan allpossible evasion maneuvers separately, and it was therefore chosen as the pathplanning method. Implementation of the safety zone constraint on multiplesurrounding vehicles, spread over multiple lanes in a Model Predictive Controlframework, together with the requirements of being able to overtake othervehicles within the prediction horizon, necessitated the development of a novel3-step algorithm. In this algorithm three Optimal Control Problems are solvedconsecutively to plan a path which positions the ego-vehicle optimally in thetargeted gap without breaching the safety zones. To be able to implement thealgorithm in real-time, a computationally expedient vehicle model is chosen andthe OCP’s are solved using state-of-the-art structure exploiting Interior Pointsolvers, generated as efficient C-code using FORCES Pro. With this setup, computationtimes of ～30[ms] are achieved for a 10[s] prediction horizon sampled at100[ms]. The algorithm is demonstrated in simulation and shown to be able tofind the point of maximum lateral intrusion, to operate on the edge of thesafety zone constraint, and to avoid a collision when it is being cut off frombehind or when the vehicle in front comes to a sudden stop. The results showthat vehicle operation by an ADS is possible in dense traffic withoutcompromising safety. By using the developed safety zones and novelty solutionframework the necessary margins can be reduced, increasing the availability ofthe lane change maneuver. Furthermore the control strategies necessary forimplementation are shown to be applicable in real-time, paving the way for implementationin real vehicles. This way the research contributes to the next step indeveloping a usable Automated Driving System.

1.  引言

2. 路径规划

3. 模型

4. 安全区

5. 研究方法

6. 实时实现

7. 研究结果

8. 结论

更多精彩文章请关注公众号：