Efficient Real-Time Pixelwise Object Class Labeling for Safe Human-Robot Collaboration in Industrial Domain
Proceedings of The 4th Workshop on Machine Learning for Interactive Systems at ICML 2015, PMLR 43:6-14, 2015.
In this paper, we use a random decision forests (RDF) classifier with a conditional random field (CRF) for pixelwise object class labeling of real-world scenes. Our ultimate goal is to develop an application which will provide safe human-robot collaboration (SHRC) and interaction (SHRI) in industrial domain. Such an application has many aspects to consider and in this work, we particularly focus on minimizing the mislabeling of human and object parts using depth measurements. This aspect will be important in modelling human/robot and object interactions in future work. Our approach is driven by three key objectives namely computational efficiency, robustness, and time efficiency (i.e. real-time). Due to the ultimate goal of reducing the risk of human-robot interventions. Our data set is depth measurements stored in depth maps. The object classes are human body-parts (head, body, upper-arm, lower-arm, hand, and legs), table, chair, plant, and storage based on industrial domain. We train an RDF classifier on the depth measurements contained in the depth maps. In this context, the output of random decision forests is a label assigned to each depth measurement. The misclassification of labels assigned to depth measurements is minimized by modeling the labeling problem on a pairwise CRF. The RDF classifier with its CRF extension (optimal predictions obtained using graph cuts extended over RDF predictions) has been evaluated for its performance for pixelwise object class segmentation. The evaluation results show that the CRF extension improves the performance measure by approximately 10.8% in F1-measure over the RDF performance measures.