<![CDATA[In general, detection of human activity is carried out to detect activity in daily life. With further development, the detection of human activity is utilized to detect suspected activity (not routine) as an early warning application. Detection of human activity in early warning applications will then be implemented in other systems. However, there are several problems in detecting human activity, among others, the presence of variations in performing an activity, the movement of transitions between activities, and the similarity of movements in different activities. Detection of human activities with existing variations can be done if utilizing a deep learning approach to conduct the training process. The deep learning method used is VGG16. VGG16 will receive input in the form of skeleton data images. The skeleton data used is obtained from the NTU RGB+D dataset. Skeleton data will be represented as 2D images by going through a process of covering the crop, converted into grayscale, resizing, and connecting 10 images into 10 channels for each sequence of activity sequences. To detect human activity is applied transfer learning on VGG16 that is changing the fully connected layer. VGG16 modification test results with skeleton data representation resulting in the highest accuracy rate of 54.59%. This level of accuracy is obtained from model testing using the same dataset as the training dataset. The VGG16 modification is still the best model based on testing with other Convolutional Neural Network models. Modification of VGG16 can classify indoor activities.]]>
