Using Human Skeleton Point Repair Methods to Improve the Accuracy of Multi-person Interactive Activity Recognition

Human motion recognition is a very important research topic in the field of computer vision. Multi-person motion is a key branch of this topic, and is also the most challenging field of computer vision. When multi-person interactive action is involved, it is easy to generate incorrect skeleton keypoint predictions, which affects the accuracy of activity recognition, since a human activity recognition model is trained to recognise the changes in skeleton points over time and space. In this paper, we propose a method of repairing the skeleton keypoints to improve the accuracy of multi-person interactive action recognition. We use the left–right symmetry of the human skeleton for the symmetry keypoint repair method, and the backbone elongation property of the human skeleton to perform the key point elongation repair method. After repairing the skeleton keypoints using the above methods we use Long Short-Term Memory to carry out training and prediction with a multi-person interactive activity recognition model. Specifically, we carry out a comparison between our proposed approach and spatial temporal graph convolutional networks, long short-term memory, and the skeleton point correction method in terms of recognition accuracy, and find that key results of the proposed approach improves the average accuracy of multi-person interactive activity recognition by at least 49.7%, 21.9%, and 8.6%, respectively.