GL-YOLO-Lite: A Novel Lightweight Fallen Person Detection Model

Area of Science:

• Computer Vision
• Artificial Intelligence
• Machine Learning

Background:

• Deep learning models for fallen person detection (FPD) struggle with global context, feature extraction, and computational demands, leading to low accuracy and slow speeds.
• Existing methods often fail to adequately utilize comprehensive contextual information, hindering robust detection in diverse scenarios.

Purpose of the Study:

• To develop a novel, lightweight deep learning model for accurate and efficient fallen person detection.
• To address limitations in current models, including poor feature representation and high computational costs.

Main Methods:

• Proposed the Global and Local You-Only-Look-Once Lite (GL-YOLO-Lite) model, integrating transformer and attention modules into YOLOv5 for enhanced contextual information utilization.
• Introduced a new stem module, rep modules with re-parameterization, and a lightweight detection head to optimize the architecture.
• Utilized a binary cross-entropy (BCE) loss function for classification and confidence calculations.
• Developed a large-scale, well-formatted Fallen Person Detection Dataset (FPDD).

Main Results:

• GL-YOLO-Lite significantly outperformed state-of-the-art models on both the FPDD and Pascal VOC datasets, achieving mean average precision (mAP) gains of 2.4-18.9 and 1.8-23.3, respectively.
• The model demonstrated real-time processing capabilities, achieving 56.82 FPS on a Titan Xp and 16.45 FPS on a HiSilicon Kirin 980.
• The enhanced feature extraction and contextual integration led to improved detection accuracy and generalization.

Conclusions:

• GL-YOLO-Lite offers a superior solution for fallen person detection, balancing high accuracy with computational efficiency.
• The model's effectiveness in real-world scenarios makes it suitable for safety-critical applications.
• The integration of global and local context through transformer and attention mechanisms is key to advancing object detection performance.