Browsing by Author "Uimonen, Mikael"
Now showing 1 - 2 of 2
- Results Per Page
- Sort Options
- Hevosen asennon estimointisyvyyskamerakuvaa hyödyntäen
Sähkötekniikan korkeakoulu | Bachelor's thesis(2020-10-30) Uimonen, Mikael - Multimodal interface for interactive cooperation with quadruped robots
Perustieteiden korkeakoulu | Master's thesis(2022-12-12) Uimonen, MikaelA variety of approaches for hand gesture recognition have been proposed, where most interest has recently been directed towards different deep learning methods. The modalities, on which these approaches are based, most commonly range from different imaging sensors to inertial measurement units (IMU) and electromyography (EMG) sensors. EMG and IMUs allow detection of gestures without being affected by the line of sight or lighting conditions. The detection algorithms are fairly well established, but their application to real world use cases is limited, apart from prostheses and exoskeletons. In this thesis, a multimodal interface for human robot interaction (HRI) is developed for quadruped robots. The interface is based on a combination of two detection algorithms; one for detecting gestures based on surface electromyography (sEMG) and IMU signals, and the other for detecting the operator using visible light and depth cameras. Multiple architectures for gesture detection are compared, where the best regression performance with offline multi-user data was achieved by a hybrid of a convolutional neural network (CNN) and a long short-term memory (LSTM), with a mean squared error (MSE) of 4.7 · 10−3 in the normalized gestures. A person-following behavior is implemented for a quadruped robot, which is controlled using the predefined gestures. The complete interface is evaluated online by one expert user two days after recording the last samples of the training data. The gesture detection system achieved an F-score of 0.95 for the gestures alone, and 0.90, when unrecognized attempts due to other technological aspects, such as disturbances in Bluetooth data transmission, are included. The system to reached online performance levels comparable to those reported for offline sessions and online sessions with real-time visual feedback. While the current interface was successfully deployed to the robot, further advances should be aimed at improving inter-subject performance and wireless communication reliability between the devices.