| Title |
User Specified Gesture Recognition Method using Corrected Machine Learning |
| Authors |
문백산(Backsan Moon) ; 김대원(Daewon Kim) |
| DOI |
https://doi.org/10.5573/ieie.2019.56.1.57 |
| Keywords |
Calibrated random forest ; Gesture recognition ; Customizing algorithm ; Machine learning |
| Abstract |
In this paper, we propose Calibrated Random Forest(CRF) classification algorithm which corrects existing Random Forest (RF) classifier to estimate user’s gesture motion more precisely in an actual environment, and a combination of several feature vectors that enables the classifier to obtain maximum information from sensor data. Most of the gestures have many individual differences, so we use calibration data obtained from actual user’s gesture to correct learned RF classifier for improving recognition accuracy. In addition, we extract and combine histogram, standard deviation and re-sampled data from the sensor data to improve the specificity of feature vectors, as a result, accuracy of the classifier was improved. We collected 30 gestures data and evaluated the accuracy of the CRF classifier and the effect of combination of feature vectors through the experiments. As a result of the comparison experiment using the combination of sensor data and feature vectors, the accuracy of 95.1% was achieved, that improved by up to 20% compared with the using only the histogram of the single sensor group. The average accuracy of CRF was 94.2% which is improved by about 48% compared to the comparison group of other classifier models. Through the research, we confirmed that it is more effective to learn by combining various sensors and features than using a single sensor data and feature vector, and the proposed CRF method classifies the individual gesture motion appropriately in an actual environments. |