基于端到端学习的视觉车道线保持方法研究

A Method of Visual Lane Keeping Based on End-to-End Learning

针对传统自主驾驶汽车复杂的车道检测、路径规划和运动控制技术问题,基于卷积神经网络设计了一种视觉车道保持系统,该方法可直接从视觉传感器中获取数据以控制车辆转向。基于消失线方法对图像进行区域剪切获取感兴趣区域,解决了训练过程需标记大量的数据集而时间成本高的问题。采用上、下采样结合及色彩空间转换方法进行数据增强,避免了数据不平衡和过拟合现象。最后,结合实际情况修正了转角损失与油门损失权重比。将预处理后的数据馈送到神经网络进行训练,通过仿真实验验证了方法的可行性,实现了基于端到端学习的DIY小车在所设置轨道上的自主驾驶。

Aiming at the complex lane detection,path planning and motion control technology of traditional autonomous driving vehicles,a visual lane keeping system is designed based on the convolutional neural network,which can directly obtain data from the vision sensor to control the vehicle steering.Find the region of interest based on the vanishing line method.This solves the problem of requiring a large amount of time to mark a large data set during training.The method of up-down sampling and color space conversion is used to enhance the data,which avoids the phenomenon of data imbalance and overfitting.Finally,the weight ratio between the Angle loss and the throttle loss is corrected according to the actual situation.The pre-processed data were fed to the neural network for training.The feasibility of the method was verified by simulation experiments,and the autonomous driving of the DIY car on the set track based on end-to-end learning was realized.