YOLO-I3D: Optimizing Inflated 3D Models for Real-Time Human Activity Recognition.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Human Activity Recognition (HAR) is crucial for surveillance and healthcare.
• Existing two-stream models like Inflated 3D (I3D) are computationally intensive, hindering real-time performance.
• The optical flow branch in I3D models significantly contributes to high computational demand.

Purpose of the Study:

• To develop a more computationally efficient two-stream model for real-time Human Activity Recognition.
• To address the limitations of high computational cost in existing HAR methods.
• To improve the performance and efficiency of I3D models for practical applications.

Main Methods:

• Proposed a lightweight motion information branch using lower-resolution RGB input to replace the optical flow component.
• Integrated YOLOv5, an efficient object detection model, to optimize the RGB branch.
• Evaluated the proposed model, termed I3D Light, on the Kinetics-400 dataset.

Main Results:

• The I3D Light model demonstrated a 4.13% accuracy improvement over the original I3D model.
• Computational cost was significantly reduced by replacing the optical flow branch.
• Integration with YOLOv5 further enhanced accuracy by 1.42%.

Conclusions:

• The proposed I3D Light model offers a more efficient solution for real-time Human Activity Recognition.
• Lightweight motion processing and object detection integration are effective strategies for improving HAR models.
• The study successfully balances accuracy and computational efficiency for practical HAR systems.