针对无人车转角输出的端对端方法

针对端对端学习过程中的数据不均衡、时间成本高、输出不够鲁棒等问题,通过数据均衡、图像尺寸变换及双边滤波对数据集进行优化,降低了卷积神经网络(CNN)模型输出的误差,此外使用固定区域的图像剪切与图像尺寸变换降低了模型训练的时间成本。分别对是否经过均衡与处理的数据集进行训练获得两种模型,首先将两种模型的输出与原始数据进行对比,此外对平均训练时间进行比较,最终在智能小车上进行了自动驾驶实验。证明所提出方法改善了端对端输出的鲁棒性、降低了模型训练的时间成本。

Research on End-to-end Learning Method for Angle Output of Autopilot Model

In order to solve the problems of data unbalance, high time cost and insufficient output robustness in the end-to-end learning process, the data set is optimized by data equalization, image size conversion and bilateral filtering. Besides these the error in output of the convolutional neural network(CNN) mode is reduced. In addition to, image clipping and image size conversion was used to reduce the time cost of model training. Two models are trained separately for the data set that whether has been equalized and processed. Firstly, the output of the two models was compared with the original data. Secondly, the average training time was compared. finally, the self-driving experiment was carried out in the smart car. It is proved that the method improves the robustness of end-to-end output and reduces the time cost of model training.