This paper discusses the usage of dierent image features and their combination in the context of estimating the motion of rigid bodies (RBM estimation). From stereo image sequences, we extract line features at local edges (coded in so called multi-modal primitives) as well as point features (by means of SIFT descriptors). All features are then matched across stereo and time, and we use these correspondences to estimate the RBM by solving the 3D-2D pose estimation problem. We test dierent feature sets on various stereo image sequences, recorded in realistic outdoor and indoor scenes. We evaluate and compare the results using line and point features as 3D-2D constraints and we discuss the qualitative advantages and disadvantages of both feature types for RBM estimation. We also demonstrate an improvement in robustness through the combination of these features on large data sets in the driver assistance and robotics domain. In particular, we report total failures of motion estimation based on only one type of feature on relevant data sets.
Lecture Notes in Computer Science, 2009, Vol 5064, p. 280-304