结合Spatial CNN的端到端自动驾驶研究  

End-to-End Automatic Driving Based on Spatial CNN

作　　者：单兆晨 黄丹丹 耿振野[1] 刘智[1] SHAN Zhao-chen;HUANG Dan-dan;GENG Zhen-ye;LIU Zhi(School of Electrical and Information Engineering,Changchun University of Science and Technology,Changchun 130022)

机构地区：[1]长春理工大学电子信息工程学院,长春130022

出　　处：《长春理工大学学报（自然科学版）》2021年第3期102-108,共7页Journal of Changchun University of Science and Technology(Natural Science Edition)

基　　金：吉林省科技厅自然科学基金项目(20200201167JC)。

摘　　要：针对当前自动驾驶中端到端深度学习算法需要庞大数据集作为训练支撑且缺少针对性的问题,基于深度迁移的思想,提出了迁移预训练VGG-16网络结合Spatial CNN网络结构的端到端自动驾驶模型。将预训练模型在ImageNet数据集上已经学习到的图像识别能力迁移至转向预测任务上,同时嵌入Spatial CNN网络结构挖掘空间特征信息。研究结果表明:在基于同等少量样本的训练后,迁移学习模型提取的特征更具有相关性,与从零开始训练的DAVE-2模型相比,预测误差率降低11.1%。在测试地图上模型预测值能很好地跟随真实值变化,说明模型能够实现高精度预测。Aiming at the problem that the current end-to-end deep learning algorithm in automatic driving needs a large data set as training support and lacks pertinent,based on the idea of deep-transfer learning,an end-to-end automatic driving model combining Spatial CNN network structure with transfer pre-training VGG-16 model was proposed.The image recognition ability that the pre-training model had learned on the ImageNet data set was transferred to the steering prediction task,and Spatial CNN network structure was embedded to mine Spatial feature information.The results show that after training based on the same small number of samples,the characteristics extracted from the transfer learning model are more correlated;and the prediction error rate is 11.1%lower than that of the DAVE-2 model trained from scratch.On the test map;the predicted value of the model can well follow the real value,which indicates that the model can achieve high precision prediction.

关 键 词：SCNN模型 自动驾驶 深度学习 迁移学习 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]