Detection of Road Risk Sources Based on Multi-Scale Lightweight Networks

Area of Science:

• Computer Vision
• Road Safety Engineering
• Machine Learning

Background:

• Manual road risk detection is inefficient and time-consuming.
• Automated detection systems are crucial for enhancing road safety.
• Existing two-stage network models face challenges in training time and deployment.

Purpose of the Study:

• To develop a novel automated approach for detecting road risk sources.
• To improve the efficiency and accuracy of road risk identification.
• To enable precise measurement of detected road risks.

Main Methods:

• Developed a Multi-Scale Lightweight Network (MSLN) using MobileNetV2 for feature extraction.
• Applied image preprocessing techniques including grayscale conversion, median filtering, and adaptive histogram equalization.
• Incorporated transfer learning to optimize training efficiency.
• Established a world-to-pixel coordinate system mapping for dimension calculation.

Main Results:

• The MSLN model demonstrated a faster convergence rate compared to traditional methods.
• Achieved enhanced precision in detecting road surfaces, potholes, and scattered objects.
• The coordinate transformation model accurately calculated the dimensions of detected road risks.

Conclusions:

• The MSLN offers a highly efficient and accurate automated solution for road risk detection.
• The proposed method addresses limitations of existing two-stage models, improving training and deployment.
• This approach significantly contributes to advancing road operation safety through intelligent detection.