Human Pose-Based Activity Recognition Approaches on Smart-Home Devices

Abstract

AbstractWith the gradual improvement of the intelligent degree of smart-home devices, its popularity is also multiplying. These devices dramatically improve the richness of available data, including a large number of indoor family life visual data. At the same time, it also has higher requirements on effectively using these data and further improving the intelligence of smart-home devices. In this paper, we mainly verify the performance of frameworks and feasibility of deploying the human activity recognition models when the computing power of edge computing devices is limited. It includes typical deep learning methods, CNN and GCN-based recognition methods, and a single activity judgment recognition method based on CTW. We also analyze the help of different data preprocessing steps in improving time efficiency and accuracy. In addition, a human activity dataset is built based on the actual home fitness equipment. The experimental results verify the feasibility and effectiveness of deploying the activity recognition model on IoT devices with limited computing capability.KeywordsPose estimationHuman activity recognitionSmart home