goldenAGER: A Personalized Feature Fusion Activity Recognition Model for Elderly

Physical and mental health are impacted as a person grows old. A Human Activity Recognition (HAR) system, which tracks a person's activity patterns and intervenes in case of an abnormal activity, could help elderly individuals to live independently. However, because of the strong intra class correlation between different activities, it is a challenging task to recognise such activities. Therefore, we proposed a personalized feature fusion algorithm, goldenAGER, which can be used to build as a model for abnormal activity recognition. In the initial stage, it extracts handcrafted HOG features and self-learned VGG-16 features to provide a rich description about the internal information of images. Then, the extracted features are provided as two different inputs to the deep neural network which are finally concatenated to classify the action type. The dataset is collected from the elderly volunteers over the age of 60 in a homogeneous environment consisting 10 classes of activities. The fusion of the features has resulted in 95% accuracy on primary dataset. The performance of the proposed model has also been tested on Microsoft Research (MSR) Action dataset giving accuracy of 93.08%. A comparison of our proposed model with the other existing models is also performed which shows that our model outperformed the existing models.

Automated summary

Physical and mental health of elderly individuals can be improved with a Human Activity Recognition (HAR) system. We proposed a personalized feature fusion algorithm, goldenAGER, which extracts handcrafted HOG features and self-learned VGG-16 features to recognize abnormal activities. Our model achieved 95% accuracy on a primary dataset and 93.08% accuracy on the Microsoft Research (MSR) Action dataset, outperforming existing models.