Research On Monocular Depth Estimation Based On Unsupervised Learning

In recent years,as the performance of graphics computing devices continues to improve,self-driving and home robots have begun to enter people's attention.The implementation of these techniques relies on accurate depth information,but the equipment for obtaining depth information is now quite expensive.In order to reduce the cost of collecting depth information,many scholars have begun to study the depth estimation algorithm.the monocular depth estimation algorithm is the closest to the actual life application,but because the monocular view provides less information,it has become a difficult problem in the field of computer vision.With the rise of deep learning,convolutional neural networks are applied to monocular depth estimation tasks,but due to the lack of datasets,supervised learning algorithms are greatly limited,and unsupervised learning algorithms receive more attention.First,this paper proposes a depth estimation framework based on unsupervised learning.The framework consists of two networks: Depth estimation network and pose estimation network.the depth estimation network is used to predict the depth map and the pose prediction network is used to predict camera motion.The framework optimizes the network by minimizing the photometric error.In order to solves the problem that the photometric error does not work in the case of illumination changes,the framework use the pixel coordinate relationship between matching points to enhance the constraint ability of the model and the robustness of the model in handling illumination changes.In addition,for problem that the photometric error can not contribute to the training of the network in the texture-less region,this paper proposed a method using the epipolar constraint to constrain the framework.Epipolar constraints are not affected by texture-less areas and can compensate for the failure of photometric errors in texture-less areas.Besides,in this paper,we introduce the constraints between non-adjacent frames to improve the performance of model.Finally,we describe the details of the training process.Besides,comparing with the result of other methods,it shows that the method we propose can effectively improve the accuracy of depth prediction and enhance the adaptability of the model to texture-less areas and illumination changes.