Improving Air-Writing Accuracy Through Data Regression and Interpolation in a Single Radar System
Seungheon Kwak, Chanul Park, Seongwook Lee · 2024
In the radar-based air-writing, the hand movement may not be completely detected depending on the transmission cycle of the radar waveform, which potentially reduces the legibility of the air-writing results. Therefore, in this paper, we propose a method of interpolating the unmeasured portions of the air-writing using polynomial regression. First, using a radar sensor, we acquire range and angle detection results for hand motion over time. Then, the trajectory of the hand motion is expressed in two-dimensional distance coordinates using the range and angle information. Subsequently, the polynomial regression is used to derive the relationship between the observation time and the change of distance coordinates in each dimension. Next, we interpolate the unmeasured portions using the derived regression model and generate the interpolated air-writing results. Finally, to verify the performance of the proposed method, the recognition accuracy of non-interpolated and interpolated air-writing results is evaluated. In other words, we evaluate the recognition accuracy when the two results are input to a convolutional neural network trained with the modified national institute of standards and technology database. The average recognition accuracy of interpolated air-writing results was 65.4% p higher than that of non-interpolated air-writing results.