Sdvo:ldso+ semantics, direct French Slam (RAL 2022)

Paper reading ：Semantic-Direct Visual Odometry

Motivation

Although the direct method SLAM More robust in textureless environments , However, due to the convexity of gray-scale image, the convexity of photometric error is only maintained in a small region , So the traditional direct vision SLAM When the tracking point has a large displacement , May fall into suboptimal local minima , The specific problem description is shown in the figure below , On the left are the gray-scale image and semantic probability image of the corresponding region , The corresponding 3D visualization in the right figure , Grayscale images preserve the details of the object , The road probability is mainly generated on the road boundary , For points on the boundary of semantic objects , The convexity of semantic probability holds in a larger region than the gray-scale image .

in addition , At present, although there are some methods that can track dynamic objects SLAM, But most of them use optimization items such as re projection error , These are all 6 A degree of freedom . This is actually not consistent with the real world , for instance , Most cars have only 3 A degree of freedom ,2 Degrees of freedom in translation ( Road plan ),1 Degrees of freedom rotating ( The normal vector of the car itself ). Therefore, it is necessary to eliminate the degrees of freedom that do not correspond to the physical meaning . The solution to this problem , The author chose to represent constraints as mechanical joints , Because the movement of mechanical joints is limited , This reduces the degrees of freedom .

Contribution

1. The proposed SDVO It is the first visual monocular that uses semantic probability to match directly SLAM System .

2. By integrating the direct alignment of semantic probabilities into LDSO in , Improved positioning performance , be better than ORB-