We present a practical bin-picking system that uses stereo vision from multiple viewpoints and the three-dimensional recognition using salient features and uncertainty evaluation. First, the salient features of objects in images are extracted. Under the guidance of the features, the two-dimensional object candidates are detected. And the feature correspondence among stereo images is made based on the global features of an object model. Second, the three-dimensional features are reconstructed using the stereo vision results. Next, with the object features and the three-dimensional measurements, the uncertainty is evaluated. Then the results from the different viewpoints are integrated to determine the optimal results for each object by using the uncertainty and the three-dimensional position of object candidates. Based on this practical three-dimensional vision, an intelligent robot system is realized to pick up objects in a bin.
