Truck model recognition for an automatic overload detection system based on the improved MMAL-Net.

Area of Science:

• Computer Vision
• Logistics and Transportation Engineering
• Road Safety Engineering

Background:

• Truck overloading significantly compromises road infrastructure integrity and traffic safety.
• Accurate detection and prevention of truck overloading are critical for sustainable logistics and public safety.
• Existing methods often lack the precision required for real-time, reliable identification of overloaded vehicles.

Purpose of the Study:

• To introduce a novel method for detecting truck overloading using advanced image recognition techniques.
• To enhance the accuracy and efficiency of identifying truck models for overloading assessment.
• To integrate truck model recognition with weighing station data for real-time overloading alerts.

Main Methods:

• Utilized an improved MMAL-Net (Multi-Modal Attention Learning Network) for truck model recognition from frontal and side images.
• Applied APPM (Attention-based Part Proposal Module) for local segmentation and part recognition in side truck images.
• Integrated the recognition system with automatic weighing station data and license plate information.

Main Results:

• The improved MMAL-Net achieved 95.03% accuracy on the Stanford Cars benchmark dataset.
• Recognition accuracy reached 85% with 20 training samples and 100% with 50 training samples on a small-scale dataset.
• The integrated system enables real-time assessment of truck overloading by combining model recognition, weight data, and license plates.

Conclusions:

• The proposed method offers a highly accurate and reliable solution for detecting truck overloading.
• Integration with existing infrastructure (weighing stations) facilitates practical implementation for enhanced road safety.
• This technology is vital for mitigating the negative impacts of truck overloading on infrastructure and public safety.