Recent research on human action recognition is largely facilitated by skeletal data, a compact representation composed of key joints of the human body that is efficiently extracted from 3D imaging sensors and that offers the merit of being robust to variations in the environment. However, leveraging the capabilities of artificial intelligence on such sensory input imposes the collection and annotation of a large volume of skeleton data, which is extremely time-consuming, troublesome and prone to subjectivity in practice. In this paper, a trade-off approach is proposed that utilizes the recent contrastive learning technique to surmount the high requirements imposed by traditional machine learning methods on labeled skeletal data while training a capable human action recognition model. Specifically, the approach is designed as a two-phase semi-supervised learning framework. In the first phase, an unsupervised learning model is trained under a contrastive learning fashion to extract high-level human action activity semantic representations from unlabeled skeletal data. The resulting pre-trained model is then fine-tuned on a small number of properly labeled data in the second phase. The overall strategy helps identify rules for using least amounts of labeled data while achieving a human action recognition model compatible with state-of-the-art performance. The framework integrates the popular graph convolutional neural networks into the proposed semi-supervised learning framework and experimentation is conducted on the large-scale human action recognition dataset, NTU-RGBD. The paper provides comprehensive comparisons between experimental results obtained with the semi-supervised learning model and with fully supervised learning models. Relative usage of labeled data is emphasized to demonstrate the potential of the proposed approach.
Previous Article in event Previous Article in session
Next Article in event
Enhancing Indoor Position Estimation Accuracy: Integration of IMU, Raw Distance Data, and Extended Kalman Filter with Comparison to Vicon Indoor Positioning System DataNext Article in session
Semi-Supervised Adaptation for Skeletal Data Based Human Action Recognition
Published: 15 November 2023 by MDPI in 10th International Electronic Conference on Sensors and Applications session Student Session
https://doi.org/10.3390/ecsa-10-16083 (registering DOI)
Keywords: human action recognition; semi-supervised learning; contrastive learning