基于卷积神经网络的道路检测方法

Road detection method based on convolutional neural networks

为提高道路检测的鲁棒性,优化高级辅助驾驶系统的性能,提出一种基于卷积神经网络的道路检测方法。采用编解码思想,用由卷积层和下采样层构成的编码器网络提取低尺度图像特征,其中下采样层是由最大池化层和卷积层构成,更好地保留图像边缘信息;用与编码器网络相对应的解码器网络将低尺度编码特征映射到原始图像尺度空间,实现像素级分类。在KITTI数据集上的实验结果验证了该方法的有效性和鲁棒性。

To improve the robustness of road detection and optimize the performance of advanced driver assistance systems （ADAS）, a road detection method based on convolutional neural networks （CNNs） was presented. Codec thought was used, the lower resolution image feature maps were learnt using an encoder network, which consisted of convolutional layers and down sample layers, and down-sample layer consisted of max-pooling layer and convolutional layer. A decoder network was used to map the lower resolution encoder feature maps to full input resolution features for pixel-wise classification. Experimental results on KITTI dataset verify the effectiveness and robustness of the proposed method.