Research And Implementation Of Depth Estimation Algorithm Based On Multi-clue Information

Depth estimation plays a fundamental role in the stereo vision tasks,including 3D reconstruction,robotics,autonomous driving,etc.In recent years,the convolutional neural network,which has developed rapidly and made remarkable achievements in compute vision,provides a new idea for the task of depth estimation.Deep-learning based depth estimation aims to predict the distance of each points by understanding the content of the image.Because of the scale-ambiguous,the key lies in the network structure and information clues.From the perspective of exploring the effectiveness of different information clues and reasonable network structure for the task of depth estimation,this paper proposes two algorithms using different clues to improve the performance of depth estimation algorithm,which includes supervised and unsupervised methods,and in the format of the ground truth contains absolute and relative depth.Firstly,this paper proposes depth estimation algorithm based on semantic information.The algorithm uses the prediction of the semantic segmentation to predict depth maps for a specific semantic area.It simplifies the complex scene by avoiding the direct regression of a wide distribution of depth values and effectively improves the accuracy of the depth prediction.Considering the pool generalization of the deep-learning based depth estimation algorithm caused by the simple scenes of a certain depth dataset,this paper also improves the generalization ability of the algorithm.A unified expression of the relative depth and multidataset training strategies are proposed,so that the different formats of depth ground truth can be supervised by the same supervision.The proposed depth estimation algorithm based on semantic information effectively improves the accuracy of the prediction and obtains a very good effect in the generalization performance.Secondly,because the current depth datasets are not ideal,and the existing unsupervised depth estimation algorithms fail to make effective use of video images,this paper creatively proposes an unsupervised depth estimation algorithm based on the temporal clue.The convolutional neural network was used to model the variation rules of the depth values of the matching points at different shooting positions.It infers depth values in target view by fitting the continuous variation of those in time sequence.The accuracy of the unsupervised depth estimation algorithm is improved effectively by introducing the temporal clue,and the whole training process is in no need of any depth ground truth.The proposed two algorithms in this paper are validated on large open datasets in depth prediction field.Experiments show that this paper effectively improves the predictions of depth estimation algorithm by integrating semantic and temporal clues respectively.