Improving Small-Scale Human Action Recognition Performance Using a 3D Heatmap Volume.

Area of Science:

• Computer Vision
• Machine Learning
• Human-Computer Interaction

Background:

• Skeleton-based human action recognition is crucial but often focuses on coarse-grained actions.
• A gap exists in deep learning for recognizing small-scale, fine-grained human actions with practical significance.

Purpose of the Study:

• To develop a novel deep learning approach for fine-grained human action recognition.
• To create a unified, general model applicable across different datasets and modalities.
• To address the challenge of joint mismatch in skeleton data.

Main Methods:

• Utilized heatmap-based pseudo videos to represent human actions.
• Employed anthropometric kinematics as prior information for feature extraction.
• Developed a pre-trained model to capture common motion features across datasets.
• Partitioned the human skeleton into five parts to facilitate information sharing and overcome joint mismatch.

Main Results:

• The pre-trained model effectively captures common motion features.
• Achieved steady and precise accuracy across various training settings.
• Demonstrated effectiveness in mitigating network overfitting.
• Outperformed state-of-the-art models when fusing joint and limb modality features.

Conclusions:

• The proposed approach offers a robust solution for fine-grained human action recognition.
• The model generalizes well across different datasets and modalities.
• Leveraging kinematic priors and skeleton partitioning enhances recognition accuracy and robustness.