Object Recognition Using Multiple Viewpoints
Hunkeler Iris, 2016
Betreuende Dozierende: Safak Korkut, Rolf Dornberger
Views: 9 - Downloads: 4
Advances in artificial intelligence and computer vision have sparked the emergence of service robots which are intended to work for, and alongside humans in a dynamic environment. One of the biggest challenges besides manipulation of objects and interaction with the environment is cognition. While vision is natural and does not require much conscious thought for most humans, it is still a complex task for computers. However, recent years have shown a lot of development in areas such as detection, recognition and categorization of objects as well as scene recognition. A well tested and often used object recognition approach is feature matching. Feature matching is built upon the extraction of distinctive features of an object to recognize it again in a different image and context. This thesis applies feature matching to the area of service robots, following the use case of a mobile object recognition robot capable of remembering and identifying specific items. A prototype using the Lego Mindstorms EV3 robotics toolkit, an Android smartphone, and the ORB feature matching algorithm provided by the computer vision library OpenCV is proposed and developed. The developed object recognition robot is capable of taking pictures of an object from multiple viewpoints. The information gathered from these viewpoints is then combined using a newly created confidence estimation based on the average Hamming distance of the 100 best matches found between two images to estimate whether the reference object is present or not.The system performs well on the test data set and manages to increase recognition precision by 11% when compared to a single image version of the same algorithm. However, the performance could not be confirmed with the evaluation data set. Analysis has shown, that the proposed approach only works well in combination with images that have strong distinctive features such as text. If such strong distinctive features are missing, the algorithm produces a lot of false positive results and thus delivers an overall unacceptably low accuracy rate.
Studiengang: Business Information Systems (Master)
Fachbereich der Arbeit: Wirtschaftsinformatik & IT-Management