The main problem is how visual odometer to estimate the camera motion, VO implementation method based on image, in accordance with the need to extract features, divided into front-end and front-end feature point method of direct method does not extract features. Based front-end feature point method, has long been considered a mainstream method of visual odometer, it is stable, not sensitive to light, dynamic objects, is more mature solutions.
Computer Vision neighborhood of researchers to design more stable than many corners of local image features, such as SIFT, SURF, ORB and so on. A feature point by the key (Key-point) and descriptors (Descriper) composed of two parts. SIFT (Scale Invariant Feature Transform Scale Invariant Feature Transform) computing capacity, ordinary PC's CPU also can not be calculated in real time SIFT features, a localization and mapping. SLAM so we rarely use this "luxury" image features. ORB (Oriented FAST and Rotated BRIEF) feature is now it seems very representative of real-time image feature. It improves the detection sub-FAST without directionality problem and using fast link speed binary description of the BRIEF sub, the entire image feature extraction greatly accelerated.
According matching points, it is estimated camera motion
1. When the camera is monocular, we only know that 2D pixel coordinates, so the problem is to estimate the motion based on two sets of 2D points. The counter electrode geometry problem to solve.
2. When the camera binocular, when the RGB-D, or we get the distance information in some way, that question is based on two sets of 3D point estimate motion. The problem is usually solved by ICP.
3. If we have 3D points and their projected position in the camera, but also to estimate the movement of the camera. The problem solved by PnP.