基于ConvLSTM双通道编码网络的夜间无人车场景预测  

Dual-Channel Encoding Network Based on ConvLSTM for Driverless Vehicle Night Scene Prediction

作　　者：李想 孙韶媛[1,2] 刘训华 顾立鹏 LI Xiang;SUN Shaoyuan;LIU Xunhua;GU Lipeng(College of Information Science and Technology,Donghua University,Shanghai 201620,China;Engineering Research Center of Digitized Textile&Fashion Technology,Ministry of Education,Donghua University,Shanghai 201620,China)

机构地区：[1]东华大学信息科学与技术学院,上海201620 [2]东华大学数字化纺织服装技术教育部工程研究中心,上海201620

出　　处：《红外技术》2020年第8期789-794,共6页Infrared Technology

基　　金：上海市科委基础研究项目(15JC1400600)。

摘　　要：为了提高夜间无人车驾驶的决策速度,减少夜间交通事故发生的概率,对无人驾驶场景预测任务进行了研究。提出了基于卷积长短时记忆的双通道编码夜间无人车场景预测网络,利用两个子网络:时间子网络提取红外视频序列的时序特征,空间子网络提取红外图像的空间特征,通过融合网络融合特征,输入到解码网络中,以实现对红外视频的未来帧预测。该网络具有端到端的优点,能够实现输入视频序列,直接输出预测帧的图像,并可以预测多帧图像。实验结果表明,该网络对夜间场景预测较准确,可以预测未来1.2 s后的图像,预测速度快,为0.02 s/帧,达到了实时性要求。The task of scene prediction is studied to improve the decision-making speed of driverless vehicles for reducing the probability of traffic accidents at night.A dual-channel encoding night scene prediction network is proposed based on a convolutional long-short term memory network.First,the temporal features of infrared video sequences and the spatial features of infrared images are extracted by the temporal and spatial sub-networks,respectively.Second,spatial-temporal features obtained by the fusion network are input into the decoding network to predict future frames of infrared video.This is an end-to-end network and can predict multiple frames.The experimental results show that the proposed network is more accurate in night scene prediction and can predict images 1.2 s in the future with a fast prediction speed of 0.02 s/frame,which fulfills the real-time requirement.

关 键 词：红外图像 场景预测 卷积长短时记忆 编码网络 

分 类 号：TN219[电子电信—物理电子学]