Abstract:

In order to use the convolutional neural network to learn to estimate the depth of a street image, the method of semantic segmentation is proposed to solve the problem of boundary blurring of the depth map. The estimated depth is generated by the left and right perspective maps to generate disparity maps for unsupervised training. The semantic segmentation layer is added in the network model, and multiple atrous convolutions parallel structure is adopted to increase the receptive field, the number of image down sampling is reduced, the information loss during down sampling is decreased, and the result is more accurate. This is also the first time in depth estimation combined with cavity convolution to increase the accuracy. By training the KITTI street dataset, in addition to increasing the detection accuracy and reducing the error rate, the objects in the renderings are made clearer, and some of the original models are retained in the renderings, it ignores the details and shows the original image more completely.

Key words: depth estimation, convolutional neural network, semantic segmentation, self-driving

摘要：

针对利用深度学习方法对街道图像进行深度估计，提出采用语义分割的方法解决深度图出现边界模糊等问题；估计深度通过左右视角图生成视差图进行无监督的训练。在网络模型中添加语义分割层，采取多个空洞卷积并行的结构增加感受野，同时减少了图像下采样的次数，降低了由于下采样带来的信息损失，使得的结果更加准确。这也是在深度估计中首次与空洞卷积相结合增加准确率。通过对KITTI街道数据集进行训练，与现有结果相比，除了增加检测准确性，降低错误率之外，使得效果图中的物体更加清晰，并且在效果图中还保留了一些原模型中被忽视掉的细节信息，将原始图像更加完整的表现出来。

关键词: 深度估计, 卷积神经网络, 语义分割, 无人驾驶