基于深度学习的夜间模式下自动驾驶场景预测  

Automatic Driving Scene Prediction Based on Deep Learning in Nighttime Mode

作　　者：阮雨 孙韶媛[1] 李佳豪 吴雪平 RUAN Yu;SUN Shaoyuan;LI Jiahao;WU Xueping(College of Information Science and Technology, Donghua University, Shanghai 201620, China)

机构地区：[1]东华大学信息科学与技术学院

出　　处：《东华大学学报（自然科学版）》2019年第5期709-714,共6页Journal of Donghua University(Natural Science)

基　　金：上海市科委基础研究资助项目(15JC1400600);国家青年自然科学基金资助项目(61603089);上海市青年科技英才扬帆计划资助项目(16YF1400100)

摘　　要：为了增强自动驾驶汽车在夜间行驶时对周围场景的理解,以便汽车或驾驶员可以及时做出相应的调整,将深度学习应用于夜视图像的场景预测。采用了一种预测编码网络来预测夜视图像的场景变化,在传统的深度卷积循环神经网络的基础上,对网络结构进行了一定的调整,将预测图像与实际图像的误差在网络中进行前向传递,不断更新预测误差来调整预测结果。试验结果表明,训练得到的场景预测模型,可以预测夜间驾驶场景0.4 s后的合理未来,且改善了对于长时间预测任务中效果不好的问题,具有良好的准确性和实时性。In order to enhance the understanding of the surrounding scenes of the automatic driving car at night so that the car or driver can make timely adjustments, a predictive coding network based on deep learning was used to predict scene changes in night vision images. Based on the traditional deep convolution-recurrent neural network, the network structure was adjusted to a certain extent. The error between the predicted image and the actual image was transmitted forward in the network, and the prediction error was constantly updated to adjust the prediction result. The experimental results show that the trained scene prediction model can predict the reasonable future of the night driving scene in 0.4 s, and improve the problem of poor effect in long-term prediction tasks. The prediction method has good accuracy and real-time performance.

关 键 词：自动驾驶 深度学习 夜视图像 场景预测 预测编码网络 

分 类 号：TN219[电子电信—物理电子学]