3D reconstruction system | L3 incremental motion recovery structure (incremental SFM)

One 、 Incremental motion recovery structure

Definition ： Simultaneously recover camera parameters through camera motion （ Internal and external parameters ） And the structure of the scene （ Coordinates of 3D points ）.

among , For ordered image matching , When matching , Just match the current image with the previous and subsequent frames , The complexity is O（n）, The pairwise matching of disordered images , The complexity is O(n^2).

The operations that can be carried out before and after are ：

The main process of incremental operation recovery structure is as follows ：

Feature detection and feature matching ：
Feature points are obtained by feature detection , Then feature matching , Then use the basic camera model to fit a basic model , Then we use this model to filter the feature points , Get more reliable feature points , Remove some matching outer points , The camera attitude is obtained by using the obtained interior points to perform the least square method .

Image connection diagram
From a group of unordered images, the image connection graph is obtained according to the matching results of feature points （ Vertex is image , The boundary is the visible area ）

Here's the picture , The more connection points are located in the central part , The fewer connection points are located outside .

Due to the low luminosity of night imaging, some details are blurred , The obtained feature matching relationship is less , So focus on the outside , But there are many matching relationships between images at night .
structure Track
Connect matching points corresponding to multiple perspectives , Build into a Track

Then perform global binding adjustment

The whole algorithm flow ：

Two 、 Several problems of motion recovery structure

Binding adjustment

• It is sensitive to the selection of initial camera pairs and the order in which cameras are added
• Repeat the bundle adjustment , Low efficiency

The scale is uncertain

Repeated structure leads to incorrect feature matching

Not enough matching feature points can be found on the non Lambert surface
When shooting an object from different angles , For some feature materials , Like mirrors , It has strong anisotropy , So from another angle , Maybe the feature of the 3D point cannot be detected , Therefore, not enough matching feature points can be found .