Vision based event classification in robotic micromanipulation

摘要

Autonomous positioning of small objects to create heterogeneous structures has great potential to advance the current micromanipulation procedures. To achieve autonomous micromanipulation, it is required to recognize the manipulation events. In this work, different classification algorithms including five common supervised learning methods are assessed for identifying states of manipulation. The classifiers are trained with data that consists of 3056 video frames and validated on 2545 videos frames. The best machine learning classifiers classified the events with 92.9 % accuracy, higher than the result of logic-based classification (88.9 %).