Abstract
The authors have been working on development of methods to detect driver’s inappropriate posture with pressure distribution sensors on the driving seat of a car. Pattern recognition techniques are used to the “image” of the pressure distribution of a driving posture. In a previous research, the authors applied Higher-order Local AutoCorrelation（HLAC）features to the recognition of a driving posture. The results of the previous experiment showed that the correct recognition rate was approximately 85%, therefore it has been necessary to develop a more reliable method. This paper proposes a novel posture recognition system that is based on SIFT（Scale-Invariant Feature Transform）method for feature extraction and the bag-of-keypoints method for classification. The SIFT method is robust to changes in size, illumination, noise and to rotation, and the bag-of-keypoints method contributes to the robustness to changes in location. Results of experiments show that the mean correct recognition rate of the proposed system is more than 99 percent if both the sensors on the seat cushion and on the backrest are available and the training is done individually for each person. The recognition rate is much higher than that of the previous study. Even if the training data include other persons’ data, the recognition rate becomes a little bit low but could be acceptably high. If one of the two sensors is not available, the recognition rate can be reduced significantly. However, the recognition rate could be acceptably high if the training is done individually for each person.
