Visual SLAM Implementation Framework

Source: Cometlabs

Basic principle:

• tracking a set of points through successive frames
• these tracks are used to triangulate the 3D positions of the points to create the map
• at the same time, using the the est point locations to calculate the pose of the camera, which could have observed them (i.e. calculate real time 3D structure of a scene from the estimated motion of the camera)

Architecture

1. Front-end
   • Abstracts sensor data into models (which are good for estimation) / Processing
   • Data association
     • Short term (feature tracking); features in consecutive sensor measurements
       • Either from sparse maps or dense-maps
     • Long term ( loop closure ; associating new measurements to older landmarks
2. Back-end
   • Performs inference on the abstracted data produced by the front end