Point-yolo: A fatigue driving detection method for edge detection devices
Revolutionizing Fatigue Detection on Edge Devices
This paper introduces Point-YOLO, a lightweight network model based on YOLOv5, designed specifically for real-time fatigue driving detection on computationally limited edge devices. Addressing the challenge of achieving high accuracy at high speeds, Point-YOLO optimizes the backbone network, introduces CA-BiFPN for enhanced feature extraction, and refines the multi-scale feature fusion module to reduce parameters while maintaining precision.
The model efficiently outputs 12 eye landmarks, enabling robust fatigue assessment. With a compact size of 2.1M parameters and a CPU speed of 43ms, Point-YOLO achieves an impressive average accuracy of 88.5%, balancing lightness with performance, making it suitable for low-power edge deployment.
Quantifiable Impact: Enhancing Road Safety with Edge AI
Point-YOLO delivers tangible improvements in critical areas of road safety and operational efficiency for logistics and transportation enterprises.
0
Average Accuracy
0
CPU Inference Time
0
Parameters
0
Potential Reduction in Fatigue-Related Crashes
By implementing Point-YOLO, enterprises can significantly reduce the incidence of fatigue-related accidents, which account for approximately 20% of all traffic crashes. The model's real-time performance on ARM-architecture embedded platforms ensures that driver fatigue is detected promptly, enabling proactive intervention and enhancing overall road safety. This translates directly into fewer accident-related costs, improved public perception, and better insurance rates.
Deep Analysis & Enterprise Applications
Select a topic to dive deeper, then explore the specific findings from the research, rebuilt as interactive, enterprise-focused modules.
Point-YOLO Network Architecture
Point-YOLO re-engineers the YOLOv5 algorithm for edge devices, focusing on lightweight design, enhanced feature extraction, and optimized multi-scale feature fusion. It streamlines the network by removing the 20x20 branch and customizing head outputs for critical eye landmark detection.
Enterprise Process Flow
Input Image (608x608)
→
Data Pre-process
→
Backbone (GLiteConv, CA-BiFPN)
→
Feature Fusion (Neck)
→
Head Module (80x80, 40x40)
→
Output (Border, Classification, 12 Eye Landmarks)
Performance Comparison on CPU Environment
Ablation studies demonstrate Point-YOLO's superior balance between model lightness, accuracy, and inference speed compared to other lightweight backbones on CPU environments.
| Method | mAp (%) | Recall (%) | NME | Size (MB) | Infer Time (ms) |
|---|---|---|---|---|---|
| YOLOv5s | 98.3 | 92.8 | 2.2 | 14 | 80 |
| MobileNetV2+PointYolo | 93.5 | 87.9 | 3.5 | 3.5 | 38 |
| ShuffleNetV2+PointYolo | 92.8 | 89.2 | 3.4 | 3.5 | 35 |
| DenseNet121+PointYolo | 96.3 | 89.4 | 2.4 | 7.98 | 56 |
| GLiteConv – PointYolo | 90.3 | 85.6 | 3.6 | 3.54 | 43 |
Real-World Performance on Android Devices
After int8 quantization, Point-YOLO successfully deployed on Android devices (Huawei Mate50Pro, Snapdragon 8Gen1), demonstrating stable performance and generalization across various scenes.
34-37
Frames Per Second (FPS) on Android
This performance is maintained even after int8 quantization, which reduced model size from 3.54MB to 2.1MB, with a slight mAp decrease from 91.9% to 88.5%.
Enhanced Robustness Through Data Augmentation
To ensure strong generalization ability, especially for varying light conditions (daytime/nighttime) and diverse driver appearances, the dataset of 9750 images (including closed-eye, open-eye, yawning, talking) underwent extensive augmentation. Techniques like zoom, rotation, shear, vertical/horizontal shift, and horizontal flip were applied to create a robust model. This proactive approach ensures Point-YOLO performs reliably in complex real-world driving environments, mitigating issues caused by glasses reflections or poor lighting, though challenging lighting still presents a limitation.
Key Takeaway for Enterprises
Comprehensive data augmentation, including diverse scenarios and transformations, is crucial for building robust AI models for real-world applications like fatigue detection, ensuring high reliability under varied conditions.
Calculate Your Enterprise's Potential Savings
Estimate the significant financial and operational benefits of integrating advanced AI-powered fatigue detection into your fleet operations.
Accelerated AI Deployment Roadmap
Our streamlined implementation process ensures rapid integration of Point-YOLO into your existing infrastructure, delivering value in weeks, not months.
Initial Consultation & Needs Assessment
Understand your fleet's specific requirements, current challenges, and integration points for the Point-YOLO system. Define success metrics.
Pilot Program & Customization
Deploy Point-YOLO in a small-scale pilot, customize detection parameters and integrate with your existing telematics or driver monitoring systems. Validate initial results.
Full-Scale Deployment & Training
Roll out Point-YOLO across your entire fleet. Provide comprehensive training for fleet managers and drivers on utilizing the new fatigue detection insights and protocols.
Performance Monitoring & Optimization
Continuously monitor system performance, gather feedback, and optimize the AI model for peak accuracy and efficiency, ensuring ongoing safety improvements.
Ready to Transform Your Fleet Safety?
Point-YOLO offers a state-of-the-art solution for real-time fatigue detection, enhancing driver safety and operational efficiency. Connect with our AI experts to explore a tailored implementation plan.
Ready to Get Started?
Book Your Free Consultation.
Let's Discuss Your AI Strategy!
Lets Discuss Your Needs
Select a Date
Sun
Mon
Tue
Wed
Thu
Fri
Sat