关于纯追踪模型的讨论

On the Pure Pursuit Model

AL VINN是美国 CMU大学为其自主车 NAVL AB设计的导航系统 ,其核心是多层感知机神经网络模型 .纯追踪 (Pure Pursuit,PP)模型在 AL VINN中起着十分重要的作用 ,即为虚拟样本产生期望的转弯方向 .但在 PP模型主要出处 Pomerleau的文章中 ,有关该模型的阐述不是很准确 ,甚至存在矛盾 ,一些重要的计算公式存在错误 .为此该文准确地阐述了 PP模型的基本思想和用法 ,区分了 PP正问题和 PP反问题 ,给出了正确的计算公式 ,计算机模拟实验验证了有关结论和计算公式的正确性 .

ALVINN(Autonomous Land Vehicle in a Neural Network) is famous navigator around the world for intelligent vehicles. It is developed by Carnegie Mellon University for NAVLAB(NAVigation LABoratory). The core of the ALVINN is the Multilayer Perceptron(MLP) neural network model. Its inputs and outputs are scene images in front of the vehicle and representations of steering directions, respectively. Training samples are acquired during human driving. But, experienced human drivers usually can keep the vehicle in the optimal path. So, the training samples can not be diverse. It results in bad generalization, i.e., once the vehicle deviates from the optimal path, steering command given by the trained MLP neural network may be not appropriate. In order to make the training samples as diverse as possible, virtual samples are created from real road scene images. Pure pursuit model plays a very important role in ALVINN, i.e., it produces desired steering directions for virtual samples. But in Pomerleau's paper, where the pure pursuit model is proposed, explanations about the model are not accurate (there is also a contradiction) and some important formulae are not correct. In our paper, basic ideas and usage about the pure pursuit model are accurately explained, forward and inverse pure pursuit problems are distinguished, and correct formulae are given. Determination of desired steering directions for virtual samples is an inverse pure pursuit problem. In order to verify related conclusions and formulae, many computer simulation experiments are carried out. A vehicle is assumed to be moving on a straight road. In ideal situations, the vehicle is at the road middle and its heading direction is coincided with the road central line. For different initial conditions, i.e., positions and heading directions of the vehicle are deviated from the ideal situations, resulted moving paths are examined. If the initial deviations of the positions and heading directions from the ideal situations are small, both the pure pursuit model in Pomerleau's paper and corrected one in the paper can quickly make the vehicle moving on the road central line. If the initial deviations are large, the pure pursuit model makes the vehicle moving outside the road, and the model corrected in the paper can still make the vehicle moving on the road central line.