Handshape Recognition Using Skeletal Data

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Accurate handshape recognition is crucial for intuitive human-computer interaction.
• Existing methods often struggle with variations in hand rotation and sensor configurations.

Purpose of the Study:

• To develop and evaluate a robust handshape recognition method using skeletal data.
• To propose a novel feature vector encoding relative finger and palm orientations.
• To assess the performance of different classifiers under challenging conditions.

Main Methods:

• A new feature vector was designed, capturing relative differences between finger pointing vectors and the palm normal.
• Multiple classifiers were trained and tested on a comprehensive dataset of 48 handshapes.
• Experiments included variations in sensor setup and significant hand rotation.
• Late fusion of individual model decisions and a comparative study on a public dataset were performed.

Main Results:

• The proposed feature vector demonstrated effectiveness in distinguishing between various handshapes.
• The method showed robustness across different sensor configurations and hand rotations.
• Comparative analysis confirmed the competitive performance against existing approaches.

Conclusions:

• The novel feature vector approach offers a promising direction for skeletal-based handshape recognition.
• The method's adaptability to different conditions makes it suitable for real-world applications.
• Further research can explore advanced fusion techniques and larger datasets.