Trajectory-Based Air-Writing Character Recognition Using Convolutional Neural Network

Writing in the air can be defined as to write digit or character in a 3D space by using a finger or marker movement. It is different from the traditional writing style. However, it has become easier to track finger and joint precisely due to the extensive improvement of sensor technologies. In this research, we proposed a trajectory-based air-writing character recognition system using a convolutional neural network (CNN). The trajectories were collected using a depth camera as a three-dimensional (3D) sequence. We used 10-fold cross-validation to validate the model. The accuracy of the proposed model was 97.29%. Also, we have collected an air-writing dataset containing 26,000 characters for training and validation, and another 4,000 for testing the model. The recognition time was 14ms per character which is fast enough to implement in a real-time system.