Research And Application Of Ranging Method Based On Monocular Depth Estimation

The monocular depth estimation technology can predict the distance information between the corresponding scene and the camera from a picture or a video,to realize the remote ranging function in the visualization technology engineering control application project to assist the power safety supervision work.However,as a method of passively obtaining distance information,the accuracy of monocular depth estimation is far less accurate than that of laser ranging technology.It is necessary to continue to improve the method to improve the prediction accuracy.In terms of applications,monocular depth estimation is currently mainly used to complete tasks such as three-dimensional reconstruction,unmanned automatic driving,and intelligent robots.The question of how to apply this intelligent and convenient ranging method to daily life and engineering projects needs to be thought about and solved.Therefore,around these two aspects,combined with the ranging needs of the project,this paper carries out the research and application of the monocular depth estimation method.In terms of improving the prediction accuracy of monocular depth estimation,we explore an improvement idea that integrates multiple attention mechanisms.Starting from this idea,this paper studies a variety of typical attention mechanisms verifies the effectiveness of numerous attention blocks in the network,and designs a set of combined blocks of two fusion multiple attention mechanisms.According to the different placements and locations of the combined blocks,this paper creates the network structure of multiple combined blocks.It determines the optimal structure of the network structure of the combined blocks placed on the scale of the decoder 1/8 and 1/4 through experiments.Therefore,two kinds of converged networks are constructed using combined blocks and optimal network structures,namely a monocular depth estimation network(GSNet)that integrates global context and spatial attention mechanism and a monocular depth estimation network(CSNet)that fuses channel and spatial attention mechanism.By comparing with mainstream networks,it can be seen that the two converged networks have better predictive performance and better processing and estimation capabilities for depth maps.In terms of application,for how to realize the ranging function of visual technology engineering control application projects,this paper proposes two application methods based on monocular depth estimation—the calculation method of safe distance and the measurement method of relative distance.In both measurement methods,calibration parameters are designed and distanced using a monocular depth estimation network that blends global context and spatial attention mechanisms to reduce the measurement error of the method.Experiments show that the application method based on monocular depth estimation is achievable and interoperable and can meet certain engineering-ranging needs.