Efficient object classification using multiple views in manufacturing environments

In this paper we present a framework for rapid object classification that uses multiple views to classify visually similar automotive parts on a conveyor belt. We have constructed a dataset of 25 different manufactured parts consisting of window pillars and guides. These parts vary in size, orientation and luster and are captured from four view points. We are able to achieve a classification rate of 97.4% using our dataset. Using an object localization approach for each view we provide a method that reduces the time it takes to build the visual vocabularies by 36.6%. Our framework shows has an improved accuracy over using single view classification and is fast enough to have practical industrial applications for fine-grained classification of very similar objects. We are able to demonstrate that ORB descriptors provide superior performance and speed over SIFT and SURF descriptors. By harnessing the speed and low computational expense of using ORB features with our framework, we are able to show that our approach has practical industrial applications in improving quality control.