Modeling structured environments by a single moving camera

We address the problem of the recovery of motion and a 3D model of environment with moving monocular vision. The vision sensor developed here is able to model structured environments in real time. Features such as corners and straight lines are tracked from image sequences. Their location and the motion of the camera are estimated with Kalman filters. Unlike in the conventional solution, these estimations are separated. There is a Kalman filter for each feature of the model. With this separation, it is simple to insert new objects into the environment model as they appear in the camera view. In addition, it is possible to use several separate modelers that update the common environment model. It is shown with several tests that motion estimation works well even with rather difficult real world scenes, but to obtain better 3D models, centers and straight lines are not discriminating enough to model such scenes.