Visual mapping using learned structural priors

摘要: We investigate a new approach to vision based mapping, in which single image structure recognition is used derive strong priors for initialisation of higher-level primitives the map. This can reduce state size and speed up building more meaningful maps. focus on plane mapping use algorithm detect estimate 3D orientation planar structures key frames, are then as initialising planes The learns relationship between such appearance from training examples offline. demonstrate context an EKF visual odometry system. Preliminary results experiments urban environments show that system able build large maps with significant at average frames rates around 60 fps whilst maintaining good trajectory estimation. suggest has considerable potential.