Intelligent predetection of projected reference markers for robot-based inspection systems

Abstract

Technical advancements in optical devices like sensors and projectors have led to tremendous innovations in manufacturing metrology, not least due to reductions in cost and the use of sophisticated image processing software. More recently, methods based on machine learning have demonstrated their high potential in meeting challenges that are difficult to overcome using conventional image processing techniques. In this context, we present an approach for the intelligent predetection of projected reference markers in robot-based inspection systems. These markers support the alignment of different sensor views and do not need to be physically attached to any parts. However, their robust detection is challenging under unfavorable lighting conditions. Hence, we introduce trained models of a cascade classifier based on both synthetic and real image data. Subsequently, we present the detection performance for different shapes and designs of markers projected onto real-world sheet metal parts as used in the automotive industry. The results demonstrate that properly trained classifiers can achieve a recall and precision of 90% and higher. The use of intelligent predetection promises more robust results in the subsequent detection of projected markers and, thus, benefits image processing in particular in geometric quality assurance applications.