Spherical Region-Based Matching of Vanishing Points in Catadioptric Images

MIS, UPJV, Amiens, Francefpascal.vasseur,cedric.demonceaux g@u-picardie.frAbstract. A large literature exists for rotation estimation using van-ishing points (VP). All these VP-based methods require to match thevanishing points in consecutive frames. It is usually considered that thismatching step is trivial. However whereas some techniques exist, theysu®er from some important limitations and might not work correctly inreal robotic applications. For example, the continuity constraint (whichaims to match the pair of VPs having the lowest angular distance) can-not handle large rotations and the line matching technique is very slowto execute and assumes accurate/stable line detection. In this paper, wepresent a fast and robust method to build the correspondences of VPs incatadioptric images. This work is strongly motivated by our research onreal-time rotation estimation for dynamic vehicles using catadioptric vi-sion. The underlying idea of the proposed method consists in matching,by histogram comparison, the spherical regions de¯ned by the VPs inthe equivalent sphere. Experiments have demonstrated the e±ciency ofthis approach in terms of speed and robustness to translation, rotation,dynamic environment and image blurring.