Research On Unsupervised Monocular Vision Based Depth Estimation Algorithm

With the rapid development of artificial intelligence,computer vision,and deep learning technology,the era of intelligence is also changing rapidly in recent years.The field of computer vision has a pivotal role in the intelligent era.Computer vision research is being revolutionized with the changes of the times which has started to change from traditional two-dimensional vision to three-dimensional vision.Monocular depth estimation is an important fulcrum of the three-dimensional world in computer vision,which is a fundamental and forward-looking work.However,current unsupervised algorithms of monocular depth estimation still have limitations in dealing with complex scenes.These algorithms are weakly to cope with weak texture scenes,occlusion of pixel projection,details deficiency of the output depth image,and redundant errors in the detailed pose information of consecutive image frames.These problems eventually lead to the obtained depth information not being able to integrally reflect the image-depth mapping relationship.Therefore,this paper conducts a relevant study to address the current issues of unsupervised monocular depth estimation.The main works are as follows:In this paper,an improved unsupervised model of monocular depth estimation is proposed to accomplish the image estimation tasks by cascading the depth estimation network and the bit-pose estimation network.Firstly,to address the problem of poor prediction in weak texture scenes,this paper uses asymmetric convolution to replace the traditional convolution for feature extraction.The asymmetric convolution can strengthen the feature extraction ability across the input images and reduce the sensitivity of the model to the weak texture phenomenon.Secondly,to address the problem of missing details in the output depth image,this paper designs a feature extraction network with a fused multi-scale perceptual field in the depth estimation network.Specifically,we use the structure of different-scale convolution and dilated convolution to increase the perceptual field of the underlying network.In this structure,the network’s ability is strengthened to fuse multi-scale detail information and feature information and ensure that the model outputs a depth image with complete details.Further,to enhance the pose information of keyframes and suppress the pose information of non-key frames,this paper designs a pose estimation network with an attention mechanism to reduce the redundant errors in pose estimation of consecutive frames and ensure the accuracy of the model output to the maximum.Finally,the effectiveness of the improved unsupervised algorithm for monocular depth estimation is verified by public datasets.In summary,the improved unsupervised model of monocular depth estimation in this paper can enhance the ability of the model to cope with complex scenes and improve the estimation accuracy of the model for depth information.Meanwhile,the proposed model effectively enhances the robustness and generalization without the premise of a large number of real depth map labels.The experimental results show that the method has a better improvement in the prediction effect of unsupervised monocular depth estimation,which provides a new idea for the research of unsupervised monocular depth estimation.