Deep Learning Based Air-Writing Recognition with the Choice of Proper Interpolation Technique.

Area of Science:

• Human-Computer Interaction
• Machine Learning
• Signal Processing

Background:

• Air-writing recognition, a form of gesture recognition, involves recognizing characters and digits written in free space.
• While traditional systems require extra sensors, smart-band adoption offers flexibility.
• A key challenge is handling signal duration variability, which traditional methods like padding and truncating address with significant data loss.

Purpose of the Study:

• To investigate the effectiveness of various interpolation techniques for standardizing air-writing signal durations.
• To develop and evaluate a novel air-writing character recognition method using a 2D-Convolutional Neural Network (2D-CNN).
• To address the data loss issue associated with traditional signal length normalization methods.

Main Methods:

• Extensive investigation of different interpolation techniques on seven public air-writing datasets.
• Development of an air-writing character recognition system utilizing a 2D-CNN model.
• Evaluation of the proposed method using both user-dependent and user-independent principles.

Main Results:

• Interpolation techniques were found to be effective in ensuring minimum data loss for time-series signals.
• The developed 2D-CNN model demonstrated superior performance in air-writing character recognition across all tested datasets.
• The proposed method significantly outperformed all state-of-the-art methods in both user-dependent and user-independent scenarios.

Conclusions:

• The study successfully developed a robust air-writing recognition method that effectively manages signal duration variability.
• Interpolation is a viable statistical technique for time-series signal processing in air-writing recognition, minimizing data loss.
• The 2D-CNN based approach offers a significant advancement in the field, achieving state-of-the-art results.