Model-Based 3-D Object Identification

摘要

The ATR technique developed in this project is based on a new non-linear poseestimator rather than on search mechanisms. Low false alarm rate performance is obtained by not forming a pose invariant detector but instead by incorporating pose dependent object information within the recognition process. The ATR is factored into a computationally intensive preparation process and a fast on-line target identification process. The approach is model-based and free of assumptions about the imaging process and object characteristics, and, can be applied to ATR and the estimation of pose parameters for articulated or multi-configuration targets from image and non-image sensor data. In this work, the initial concept of the pose estimator for 1 DOF (degree-of-freedom) problems was developed into a system for N DOF whole and partially obscured target pose indexing and recognition. Performance was demonstrated at the level of filter bank implementations for 1 DOF problems at 1/17 the computational cost for unobscured targets and false alarm rates orders of magnitude better than that of the filter bank approach for obscured targets. The computational savings further increase with N for N DOF problems. The report contains ROC curves obtained from tests using the public MSTAR data set.