Related Experiment Video
Updated: Jan 9, 2026
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
996
Heatmap Pooling Network for Action Recognition From RGB Videos
IEEE Transactions on Pattern Analysis and Machine Intelligence
|December 5, 2025
Summary
This study introduces HP-Net, a novel heatmap pooling network for human action recognition (HAR) in videos. HP-Net extracts robust features, outperforming existing methods and enabling more accurate action identification.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Machine Learning
Background:
- Human action recognition (HAR) in videos is crucial but faces challenges with RGB data, including redundancy, noise, and high storage costs.
- Existing deep feature extraction methods struggle to efficiently utilize rich information present in videos.
Purpose of the Study:
- To propose a novel heatmap pooling network (HP-Net) for robust and concise feature extraction in human action recognition.
- To enhance action recognition accuracy by integrating extracted features with multimodal data.
Main Methods:
- Developed a heatmap pooling network (HP-Net) with a feedback pooling module for information-rich, robust, and concise human body feature extraction.
- Designed spatial-motion co-learning and text refinement modulation modules for multimodal data integration.
Main Results:
- The extracted pooled features from HP-Net showed significant performance advantages over traditional pose and heatmap features.
- HP-Net consistently outperformed existing human action recognition methods across multiple benchmark datasets (NTU RGB+D 60/120, Toyota-Smarthome, UAV-Human).
Conclusions:
- HP-Net effectively addresses the limitations of existing HAR methods by providing superior feature extraction and integration capabilities.
- The proposed approach offers a more robust and efficient solution for human action recognition from videos.
