基于分段学习模型的自动驾驶行为决策算法研究

在具有车道线的特定自动驾驶场景中,针对目前端到端的行为决策算法直接输入原始图像进行决策导致的网络模型迁移性差、预测精度欠佳、泛化能力不足等问题,提出一种基于分段学习模型的车辆自动驾驶行为决策算法。首先,基于GoogLeNet建立一种端到端的车道线检测网络模型,并引入车道中心线作为决策的重要线索提高算法的迁移能力,同时利用YOLOv3目标检测模型对本车道内前方最近障碍物进行位置检测;而后,经几何测量模型将两者检测结果转换成环境状态信息向量为决策做支撑;最后,构建基于长短期记忆(LSTM)网络的驾驶行为决策模型,根据编码的历史状态信息刻画出动态环境中车辆的运动模式,并结合当前时刻的状态推理得到驾驶行为参量。使用建立的真实驾驶场景数据集对模型分别进行训练、验证与测试,离线测试结果显示车道线检测模型的检测位置误差小于1.3%,车道内前方障碍物检测模型的检测精度达98%以上,驾驶行为决策网络模型表征预测优度的决定系数 大于0.7。为进一步验证算法的有效性,搭建了Simulink/PreScan联合仿真平台,多种工况下的仿真验证试验中多个评价指标均达到工程精度要求,实车测试的试验结果也表明该算法可实现复杂驾驶场景下平稳、准确无偏航的预测效果并满足实时性要求,且与传统端到端模式的算法相比,具有更好的迁移性和泛化能力。

Decision Algorithm for Autonomous Driving Behavior Based on Piecewise Learning Model

With regard to specific autonomous driving scenes involving lane lines, to overcome challenges pertaining to poor network model migration, poor prediction accuracy, and the insufficient generalization ability of existing end-to-end behavior decision algorithms (where the original image is directly used for decision making), this paper proposes an automatic decision algorithm for driving behavior based on a piecewise learning model. First, an end-to-end lane-detection network was established based on GoogLeNet, and the lane centerline was introduced as an important factor for improving the migration ability of the algorithm. The position of the nearest obstacle in front of the vehicle and in the same lane was detected using the YOLOv3 target detection model. Subsequently, the detection results were transformed into environmental state information vectors through a geometric measurement model to support the decision making. Lastly, a driving behavior decision model based on the long short-term memory network was constructed. In this model, the motion mode of the vehicle in a dynamic environment is described according to encoded historical state information, and the driving behavior parameters are deduced by incorporating the current state. This model was trained, verified, and tested using a real driving scene dataset that was built by the authors. Offline test results indicate that the error in the position detection of the lane-detection model is less than 1.3%; the accuracy of the model for detecting obstacles in front of the vehicle exceeds 98%; and the decision coefficient, R², which characterizes the goodness of the predictions, exceeds 0.7 for the driving behavior decision model. A combined Simulink/PreScan simulation platform was built to validate the effectiveness of the algorithm. Results of these simulation experiments indicate that, under various working conditions, multiple evaluation criteria meet engineering accuracy requirements. The results of an actual vehicle test also indicate that the algorithm achieves smooth and accurate predictions without deviations even under complex driving scenarios, while also meeting the real-time requirements. Compared with traditional end-to-end algorithms, the proposed algorithm features better migration and generalization capabilities.