Human Activity Detection
U-ActionNet is a unified multi-stage neural network framework for human action recognition from aerial videos that blends spatial and temporal streams. It is designed to capture long-term motion, appearance and behavior patterns effectively from aerial videos with complex human-object-human interactions.
🧠 Motivation
Aerial action recognition is particularly challenging due to:
- Complex object manipulation under occlusion and motion
- Diverse temporal lengths and scene compositions
- Multiple actors in the frame in non-trivial actions
Previous methods either:
- Focused on short-term features (e.g., optical flow, CNNs)
- Or lacked unified temporal modeling across long horizons
Flow of Human Action Recognition (HAR) framework employing Unmanned Aerial Vehicle (UAV) technology in surveillance systems
U-ActionNet addresses this gap by introducing:
- Multi-stage feature extraction with region of interest extraction using YOLO
- Fusion of spectral and temporal features to understand motion patterns
- Modular design to handle both short-term and long-term dependencies
- Edge device compatible low compute lightweight version
U-ActionNet architecture incorporates an Object Localization Block (green), m-C3D Block for feature extraction (red), Fourier Substance Separation Block (blue), Fourier Self-Attention Block (blue), and two Dense Layers (yellow)
⚙️ Key Features
- Multi-Stage Transformer Backbone:
- Region of Interest extraction using YOLO
- Feature Extraction using modified C3D module
- Fourier Based Action Recognition:
- Fourier Substance Separation module isolates dynamic action rich regions from static regions
- Fourier Self-Attention captures and memorises context from the aerial videos
- Compatibility:
- For the serverside heavy duty serverside model
- For edge device, key frame based lightweight model
📝 Citation
@article{chowdhury2024u,
title={U-ActionNet: Dual-pathway fourier networks with region-of-interest module for efficient action recognition in UAV surveillance},
author={Chowdhury, Abdul Monaf and Imran, Ahsan and Hasan, Md Mehedi and Ahmed, Riad and Azad, Akm and Alyami, Salem A},
journal={IEEE Access},
year={2024},
publisher={IEEE}
}