Construction Method of Depth Estimation Model Based on Multi-Vision Task Collaboration

Abstract

The present invention provides a method for constructing a multi-visual task collaborative depth estimation model, which includes the following specific steps: fast scene depth estimation model construction under the constraints of stereo vision; model optimization for parallax geometry and knowledge prior collaboration; and the goal of joint semantic features Depth refinement: Construct a semantic segmentation module similar to depth estimation that is optimized step by step from rough to fine, forming a symmetrical structure shared by the feature layer, and then using different network features at the same stage to obtain semantic geometric information through the parallax acquisition network The disparity map; to further achieve the goal of obstacle target refinement. The present invention embeds multi-scale, knowledge prior and visual semantics into the depth estimation model, and through the multi-task cooperative sharing learning mode, deeply approaches the essence of human perception, and improves the depth estimation accuracy of obstacles.

Description

technical field [0001] The invention relates to the technical field of electronic walking assistance equipment, in particular to a method for constructing a depth estimation model of multi-visual task cooperation in the electronic walking assistance equipment. Background technique [0002] According to the latest statistics from the World Health Organization, there are approximately 285 million visually impaired people worldwide, and only China has 20 million low-vision and blind people. Daily travel is the biggest problem faced by visually impaired people in their daily lives. Today, with the rapid development of technology and the Internet, they are more eager than ordinary people to enjoy the convenience brought by artificial intelligence. Therefore, how to benefit the visually impaired and extend their vision to perceive the surrounding environment is an important research topic. Traditional blind-guiding assistance technologies and tools have relatively large limitatio...

AI Technical Summary

Problems solved by technology

Learning-based methods can combine local context and prior knowledge to improve the accuracy of depth estimation in ill-posed areas, but such methods have a strong dependence on the dataset, which affects the scene generalization ability of the model, and more parameters are very difficult. Difficult to use on energy or memory constrained ETA devices

Images