Depth estimation from monocular image plays an essential role in artificial intelligence, which is one of the important ways for sensing the operating environment in automatic-driving system or advanced driving assistant system. The most recent approaches have gained significant improvement for depth prediction based on convolutional neural networks (CNNs). In this paper, a novel framework of CNNs is proposed for monocular depth estimation based on deep ordinal regression network (DORN) and a U-net structure. The new model is trained, verified in process and tested on 5000 images from a simulation experiment platform provide by “Grand Theft Auto”. To eliminate or at least largely reduce the impact from ground truth with no depth values, three different training strategies were employed for network optimization. We developed an effective weighted training strategy for depth prediction to improve the estimation accuracy. The comparison of evaluations over our results and DORN demonstrated the effectiveness of our method. The results showed that the proposed method achieved state-of-the-art performances.
An Improved Convolutional Neural Network for Monocular Depth Estimation
Lect. Notes Electrical Eng.
2020-03-24
9 pages
Article/Chapter (Book)
Electronic Resource
English