발간년도 : [2021]
| 논문정보 |
|
| 논문명(한글) |
[Vol.16, No.3] Improved Motion Recognition Based on Convolution Neural Network for High Accuracy |
|
| 논문투고자 |
Sangmin Suh |
|
| 논문내용 |
Motion recognition (MR) is to understand and analyze the movements of things. This technique can categorize the movements into several activities and enables us understand the movements. Among the MR, human activity recognition (HAR) is to recognize what a person is doing. HAR has various applications such as rehabilitation engineering, health care, human machine interactive, security based on vision. HAR can be achieved with the help of sensor data or camera based vision information. The vision information results from monocular/binocular cameras, whereas the sensor data in general are from wearable sensors or sensors embedded in smartphones. The measured data are analyzed to carry out HAR. The data analysis methods include machine learning and neural networks. In the neural network based data analysis, the introduced time series data are converted into images, and the images are used for training of neural networks. In this paper, a new accuracy enhanced method is proposed in convolution neural network based HAR. If two images have their unique features, person may recognize the difference between the two images. Motivated from this intuition, this paper suggests a new method to add the unique feature, resulting in more precise accuracy. In this paper, firstly an optimal image conversion method is found via base line tests and secondly a reasonable neural network without overfitting is designed by trimming hyper parameters. From the experimental results, it is verified that the proposed method is valid and effective. |
|
| 첨부논문 |
|
|
|
|
|
자료실 > 논문지 
