主催: 一般社団法人 日本機械学会
会議名: ロボティクス・メカトロニクス 講演会2017
開催日: 2017/05/10 - 2017/05/13
We present a method for performing visual SLAM from a video sequence acquired by active scope cameras (ASCs). To explore the inside of debris of a house etc. ASCs go into a narrow space, which sometimes makes the camera very close to objects in the environment. This makes it hard to employ standard visual SLAM methods, since the acquired images will vary significantly between frames, incurring difficutles of tracking features stably. The employment of a high-speed camera will mitigate this difficulty by making image changes effectively small, but also make it difficult to perform visual SLAM in real-time. To cope with this, we propose a method that decimates unnecessary frames in an adaptive manner. Experiments including those conducted in the experimental field of ImPACT TRC show the effectiveness of our approach.