自适应神经网络控制在车辆自主驾驶中的应用

Application of Adaptive Neural Network Control in Autonomous Driving of Vehicles

随着人工智能和深度学习的快速发展,确保在复杂的现实环境中无缝、安全地导航是目前自动驾驶面临的主要问题。面对多变的交通环境,这是将复杂的数学公式转化为精确的物理步骤。文章的目标是通过在真实模拟中开发和测试复杂的控制系统和深度学习技术,提高自动驾驶汽车在各种驾驶场景下的安全性和有效性。这项研究涉及AirSim模拟器的使用,用于训练神经网络独立驾驶汽车。该研究首先使用AirSim利用虚幻引擎中可用的环境,并利用神经网络操纵汽车训练它。第二部分在此基础上研究了数据源选择对神经网络性能的影响。比例-积分-微分(PID)控制器和模型预测控制(MPC)用于收集数据并在其上训练设计的模型。该报告还探讨了遇到的困难,例如模型选择和作为性能指标的均方误差的限制,为未来在真实世界应用和更复杂的控制场景中的研究奠定了基础。本研究中将使用MPC和PID控制器收集数据的结果,然后将经过训练的模型应用于该数据集。通过验证数据集的图像输出输出,以测试该模型在自动驾驶测试中的泛化能力。

With the rapid advancement of artificial intelligence and deep learning,ensuring seamless and secure navigation within complex real-world environments has emerged as a primary challenge in the field of autonomous driving.This challenge entails the translation of intricate mathematical formulations into precise physical maneuvers amidst the dynamic nature of traffic scenarios.The objective of this paper is to enhance the safety and efficacy of autonomous vehicles across a variety of driving situations through the development and testing of sophisticated control systems and deep learning techniques within realistic simulations.This research involves the utilization of the AirSim simulator for training neural networks to autonomously operate vehicles.The initial phase of this study employs AirSim to leverage the environments available within the Unreal Engine,training the vehicles via neural networks to navigate these scenarios.The subsequent part of the research examines the impact of data source selection on the performance of neural networks.Proportional-Integral-Derivative(PID)controllers and Model Predictive Control(MPC)are employed for data collection,which is then used to train the designed models.The report also delves into the challenges encountered,such as model selection and the limitations of Mean Squared Error as a performance metric,laying the groundwork for future research in real-world applications and more complex control scenarios.In this study,the outcomes of data collection using MPC and PID controllers are presented,followed by the application of the trained models to this dataset.The generalization capability of the model in autonomous driving tests is verified through the image output of the validation dataset.