Estimate Depth Information from Monocular Infrared Images Based on Deep Learning

In the field of automatic driving, vision-based driving assistance system(V-DAS) is a research direction. Aiming at the demand of V-DAS for low visibility environment perception , we propose a depth estimation method based on monocular infrared images. Our method uses an end-to-end self-supervised deep learning framework, and uses the reprojection relationship between monocular video frames to construct the loss function, which makes it easy to collect dataset. In order to mitigate the influence of infrared noise, we adopt the minimum reprojection loss method. To ensure the neural network can learn infrared features better, we use transfer learning from RGB image dataset first. Qualitative experiments on FLIR datasets show that the proposed method can obtain pixel-lev el dense depth from monocular infrared images. Experiments on real roads show that the proposed method can effectively perceive the depth information of the target in the night, and the absolute error is 13.2% within 15m. This can meet the requirements of collision avoidance in most emergency situations during automatic driving.