| Summary: | 碩士 === 國立臺北科技大學 === 電機工程系 === 107 === In recent years, activity recognition has received much attention. For example, recognizing daily activities of elderly people can detect abnormal activity of elder people living alone. This study aims to recognize daily activities of people by employing RGB activity images and deep convolutional neural network (DCNN).
The human activity database collected by Cornell University which captured by Kinect device is employed to train the DCNN model and activities classification. The database contains 12 kinds of labeled activity images extracted from activity videos and the number of image samples are 68928. Firstly, in the pre-processing stage, activity images (240*320 pixel) are downsampled to 64*64, 32*32 and 16*16 pixels for improving the computational efficiency of the model and reducing the privacy invasiveness. In this study, two DCNN models (DCNN-1 and DCNN-2) are used for activity image classification via 4-fold cross validation. Experimental results show that DCNN-2(64*64 pixels) the precision, recall, specificity, F_1-score, and accuracy of 98.8%, 98.9%, 99.9%, 98.8%, and 99.9%, respectively, are achieved with 4-fold cross validation while DCNN-2 is employed which are superior to other existing systems, indicating the application potential of our work.
|
|---|
