Statistical feature selection model for robust 3D object recognition

This paper presents the feature selection with statistic modeling in real environment for 3D object recognition and pose estimation. For robust object recognition and pose estimation in various environments, we attempt using various features (SIFT, line, and color). However, each feature's reliability changes as environment changes such as illumination, occlusion, and distance. We estimate the changes of features in different environments to make reasonable feature selection using following methods. We predict expected feature quantity by combing detection probability (statistical model) of each feature and idle feature quantity (object model) of an object that we can see current viewpoint. Moreover, we calculate each feature's reliability by combining utility function to decide if expected features are valid in object recognition and pose estimation process. Based on the final probability, we decide the optimal feature. Selecting the optimal feature in environmental change enables fusion and filtering. We can recognize objet and estimate pose under severe environments. Moreover, there is calculation benefit as does not use features with low reliability. Our method verified performance of algorithm through real environment experiments.