Assessing the computational effort for structural 3D vehicle recognition


Two alternatives for the detection of vehicles in oblique view image sequences are discussed. Both methodes make use of structural decomposition of 3D models in part-of hierarchies and concretizations. Also the production system shell is shared. The interaction of the productions is depicted by production nets. The first approach uses direct 3D matching of model and measured data. 3D faces are inferred from 2D image features by means of a stereo production. In contrast to that the second approach projects all plausible settings of model parts into a 2D image space. The matching against measured image features is then done by a geometric hashing technique. The same geometric 3D Model and part-of hierarchy of a truck is used as example to explain both approaches. Tests on the same visual spectrum domain image sequence are reported. Pros and contras for both methodes are discussed in mutual comparance as well as with respect to other published work. Recommendations depend on the details of the task and geometric setting.
Michaelsen E, Stilla U (2000) Assessing the computational effort for structural 3D vehicle recognition. In: Ferri FJ, Inesta JM, Amin A, Pudil P (eds) Advances in Pattern Recognition: Joint IAPR International Workshops SSPR 2000 and SPR 2000. Berlin: Springer, 357-366
[ Stilla.de/pub ]