基于深度学习的连续帧车道线检测网络

Continuous frame lane line detection network based on deep learning

为了改善单帧图像检测复杂背景中车道线性能较差问题,例如车道线受到阴影影响、污渍污损或人车遮挡等情况时性能较差的问题。本文提出了一种基于连续帧的车道线检测网络实现了卷积神经网络(C onvolutional Neural Network,CNN)和长短期记忆网络LSTM(Long Short-Term Memory, LSTM)的融合。首先,编码器CNN对连续帧进行特征提取,生成多尺度特征映射;其次,输入对应的双层ConvLSTM网络,捕获连续帧的时空信息;最后,捕获的时空信息在解码器CNN中进行多尺度特征融合,产生车道线预测的分割图。实验结果表明所提网络的准确率、召回率和F1值较高,分别达到了 85.8%、96.1%和90.0%,总体上F1相对于原始CNN网络提高了约4%,在某些复杂路况下F1的提升在10%以上。与其它网络相比本文提出的网络具有较高的准确率、召回率和F1值,同时运行时间并没有大幅增加实时性得到保障。

In order to improve the performance of single-frame image detection of lane lines in dealing with complex road conditions,such as lane lines affected by shadows,stains,or occlusion by people and vehicles,the performance of the lane line is poor.A lane line detection network based on continuous frames is proposed,which realizes the fusion of Convolutional Neural Network(CNN)and Long Short-Term Memory(LSTM).First,the encoder CNN performs feature extraction on continuous frames to generate a multi-scale feature map,and then enters the corresponding double-layer ConvLSTM network to capture the spatiotemporal information of the continuous frames.Finally,the captured spatiotemporal information is feature fused in the decoder CNN to produce lane line prediction.Segmentation diagram.Experimental results show that the accuracy,recall,and F1 value of the proposed network are high,reaching 85.8%,96.1%,and 90.0%respectively.In general,F1 is increased by about 4%compared to the original CNN network.F1's improvement in road conditions is more than 10%.Compared with other networks,the proposed network has a higher accuracy rate,recall rate and F1 value.At the same time,the running time has not increased significantly,and the real-time performance is guaranteed.