Depth Estimation System Using A Monocular Camera

In recent years,with the development of computer software algorithm and the breakthrough of hardware computing power,deep learning technology has made great progress.People are using deep learning technology to build the new era of artificial intelligence.An important area of artificial intelligence research is computer vision.Deep learning technology has successfully solved many difficult problems in 2D vision,and now the research focus of computer vision is gradually changing from 2D vision to 3D vision.Monocular depth estimation is a fundamental yet challenging task in 3D vision.Monocular depth estimation has a broad development prospect and many applications including unmanned vehicle,mobile robot,augmented reality(AR),3D reconstruction and so on.At present,existing monocular depth estimation methods have many limitations.Traditional methods(SFM and SLAM)leverage camera motion to estimate camera pose,and then use triangulation to restore scene depth,but the obtained depth map is usually sparse and semi-dense.However,although the dense depth map can be obtained by using convolutional neural network for monocular depth estimation,it is still in the development stage and not mature enough,and its accuracy still needs to be improved.In this paper,a fast and accurate monocular depth estimation system is designed.The whole system is mainly composed of two parts: The first part is the DeNet network.A59-layer convolutional neural network is designed for the prediction of depth map and uncertainty distribution map.Each pixel in an uncertainty map indicates the error variance of the corresponding depth estimate;The second part is the inter-frame information fusion part.Depth estimates and uncertainties of previous frames are propagated to the current frame based on the tracked camera pose,yielding multiple depth/uncertainty hypotheses for the current frame which are then fused in a Bayesian inference framework for greater accuracy and robustness.We test our system on many public datasets,the results show that our overall system's performance is significantly improved compared to other approaches.In addition,based on the monocular depth estimation system proposed in this paper,we also made a demo for 3D dense construction,and the construction effect is better than the traditional method.