Skip to main content Skip to navigation

Few-Shot Continual Learning for Activity Recognition in Classroom Surveillance Images

Project Overview

The document explores the innovative use of generative AI in education, particularly through the development of the ARIC dataset and a few-shot continual learning approach aimed at activity recognition within educational environments. It emphasizes the importance of accurately recognizing a range of activities, both teaching and non-teaching, from classroom surveillance images, while tackling challenges like class imbalance and the high similarity between different activities. The proposed method integrates supervised contrastive learning with an adaptive covariance classifier, significantly enhancing the model's performance in real-world classroom scenarios. The findings suggest that this approach not only improves activity recognition accuracy but also offers insights into classroom dynamics, thus contributing to a better understanding of educational environments and potentially informing teaching strategies and classroom management practices.

Key Applications

Few-Shot Continual Learning for Activity Recognition in Classroom Surveillance Images

Context: Educational context focusing on real classroom settings, targeting educators and researchers in AI and education.

Implementation: Developed a dataset (ARIC) and a few-shot continual learning method that integrates supervised contrastive learning and adaptive covariance classifier.

Outcomes: Enhanced model's generalization ability, improved accuracy in activity recognition, and better handling of class imbalance and high similarity among activities.

Challenges: Class imbalance, high similarity between activity types, privacy concerns, and continuous occurrence of non-instructional activities.

Implementation Barriers

Technical Barrier

Challenges in recognizing activities due to class imbalance and high similarity between activities.

Proposed Solutions: Utilizing few-shot continual learning methods to enhance the model's ability to learn from limited samples.

Privacy Barrier

Need to protect the privacy of individuals in classroom surveillance images while recognizing activities.

Proposed Solutions: Using shallow layers of pre-trained models to convert images into feature data, avoiding the release of original images.

Project Team

Yilei Qian

Researcher

Kanglei Geng

Researcher

Kailong Chen

Researcher

Shaoxu Cheng

Researcher

Linfeng Xu

Researcher

Hongliang Li

Researcher

Fanman Meng

Researcher

Qingbo Wu

Researcher

Contact Information

For information about the paper, please contact the authors.

Authors: Yilei Qian, Kanglei Geng, Kailong Chen, Shaoxu Cheng, Linfeng Xu, Hongliang Li, Fanman Meng, Qingbo Wu

Source Publication: View Original PaperLink opens in a new window

Project Contact: Dr. Jianhua Yang

LLM Model Version: gpt-4o-mini-2024-07-18

Analysis Provider: Openai

Let us know you agree to cookies