Model-based Object Localization and Pose Estimation Method Robust Against the Stereo Miscorrespondence

Abstract

The miscorrespondence in stereo image analysis, which is caused by occlusion among images with failure in edge detection, often occurs in real factory environment, and this seriously disturbs the object localization and pose estimation. This work shows that, even under such conditions, the location and attitude of target object can precisely be measured, based on the three base-line trinocular stereo image analysis, using a “model-based verification” method, i.e., a model-based object recognition method including a multi-modal optimization algorithm. This method is suitable for real applications which need object localization and pose estimation, like a bin picking of parts randomly placed on factory automation.