| Depth estimation is indispensable in the application scenarios of autonomous interaction between computer and environment,such as automatic driving and service robot applications.Monocular depth estimation based on deep learning,depth information can be obtained only from a single RGB image,which has been widely concerned and researched.In recent years,monocular depth estimation based on deep learning has made rapid development,but there are still some technical indicators such as depth estimation range,accuracy and resolution,and the efficiency and environmental applicability of the algorithm cannot fully meet the requirements of different applications and scenarios.These problems always affect the wide application of this technology in practical scenarios.Among the two paradigms of unsupervised learning monocular depth estimation,the unsupervised monocular depth estimation based on time sequence constraints only requires video image sequences collected by monocular camera,which is simpler and has lower threshold for universal research,but theoretically it has stronger generalization and application value.The work of this thesis is mainly aimed at the problems of monocular depth estimation based on time sequence constraints due to the existence of data contrary to the luminosity consistency hypothesis and scene static hypothesis,depth prediction lacks semantic information,and edge depth accuracy is not high in continuous video image sequences.A monocular depth estimation framework based on image segmentation guidance is proposed.The specific work mainly includes:1.In view of the lack of semantic information and the low accuracy of edge depth caused by scene and visual occlusion,an edge consistency module is designed to take into account the edge consistency constraint and structure constraint.More accurate segmentation edges in semantic segmentation images are used to force constraint depth estimation edges to enhance the semantic expression ability of depth estimation network.The experimental results show that compared with the baseline algorithm,the performance index of the network with the edge consistency module is significantly improved,and the prediction expression ability of the depth is further improved.To some extent,the current unsupervised monocular depth estimation edge is not effective.2.Aiming at the influence of dynamic objects on the depth estimation model in the scene,a dynamic object detection algorithm based on semantic segmentation is proposed,and a monmesh depth estimation algorithm framework based on semantic segmentation dynamic object detection is constructed.The semantic segmentation graph of adjacent frames in the image sequence is used to reconstruct the segmentation image by using the multi-view image reconstruction method.By calculating the intersection ratio between the true semantic segmentation graph and the reconstructed semantic segmentation graph,the dynamic object is judged and a dynamic mask is formed.The dynamic mask is used to suppress the pollution of the photometric projection error caused by the dynamic object.Experiments show that the proposed method is effective and the depth estimation performance of the network with dynamic object detection module algorithm is significantly improved compared with the baseline network.3.Combining the two methods proposed in this thesis,a monocular depth estimation algorithm framework based on image segmentation guidance is constructed.The working condition of the overall algorithm framework was tested through the module ablation experiment.The influence of the dynamic mask on the proposed loss function was analyzed through the dynamic mask coverage experiment,and the network loss function was adjusted to the best collocation.After experimental verification on general data set,the prediction and expression ability of the network for depth has been significantly improved,and each index exceeds the mainstream algorithms in the industry. |
