Three dimensional object recognition using a complex autoregressive model

Complex Partial Correlation (CPARCOR) features, derived from an autoregressive model, are known to provide exceptional position, scale, and rotation invariant (PSRI) properties for planar two dimensional (2-D) object recognition. Although autoregressive models have been successfully applied to numerous spatio-temporal recognition tasks, the effects of out-of-plane image rotations and known levels of occlusion have not been considered. This study investigates applications of the CPARCOR model to a five class problem of nonplanar 2-D views of 3-D objects. Recognition (based on CPARCOR features) of both single and multiple frames of imagery is performed using the hold-one-out error estimation method on a 1-Nearest Neighbor classifier. Direct comparisons to recognition based on low frequency Fourier magnitude features are made. Additionally, the effects of known levels of occlusion on the classification rate was examined using occluded nonplanar views and a template classifier. Results indicate that the CPARCOR model parameters provide useful shape-features for recognition of out-of-plane rotations. Displaying exceptional PSRI properties, the features are shown capable of classification by simple nonadaptive recognition schemes. The advantage of classification by a multiple-look technique over the traditional single-look method is dearly demonstrated. Feature space crowding is noted as the cause of unusual recognition rates for occluded-view tests. Although general trends are noted, optimal model order and selection of CPARCOR versus Fourier features are considered application dependent