Scholarlink

<p>Pavement surface conditions and weather changes significantly impact pedestrian safety in urban environments. Current pavement research primarily focuses on vehicle-centric safety indices and single-modality assessments. This study contributes new knowledge by introducing the Pedestrian Slipperiness Index (PSI). This multimodal human-centric metric integrates road surface imagery and real-time weather streams to quantify pedestrian fall risk. An ensembled convolution neural network (CNN) model combining ResNet18, EfficientNet-B0, and MobileNetV2 classifier is trained on Road Surface Classification Dataset (RSCD) dataset to predict a real-time pavement surface from 16 pavement surface classes (e.g., Wet Asphalt Severe, Dry Concrete Smooth), achieving 90% test accuracy and a macro F1-score of 88%. The image-based PSI (psi_image) for the predicted pavement surface class was effectively fused with the real-time weather-based PSI (psi_weather) using a Multi-Layer Perceptron (MLP) regressor. This regressor was trained on the Cartesian product of image and weather features, resulting in an impressive R² value of 0.999. The proposed multimodal solution combines Artificial Intelligence (AI)-based surface condition analysis and weather-based risk modeling to provide context-aware, real-time pedestrian safety, which enriches pavement research. This work offers a scalable method for proactive pavement maintenance and urban safety planning by changing the paradigm from vehicle-oriented indices to pedestrian-centric risk estimation.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Predicting Pedestrian Fall Risk Through Deep Feature Fusion of Pavement Surface Imagery and Weather Streams

  • 期刊论文

  • 发表日期：

    2026年3月8日

查看全文
  • Priti Chakurkar,
  • Deepali Vora

摘要

Pavement surface conditions and weather changes significantly impact pedestrian safety in urban environments. Current pavement research primarily focuses on vehicle-centric safety indices and single-modality assessments. This study contributes new knowledge by introducing the Pedestrian Slipperiness Index (PSI). This multimodal human-centric metric integrates road surface imagery and real-time weather streams to quantify pedestrian fall risk. An ensembled convolution neural network (CNN) model combining ResNet18, EfficientNet-B0, and MobileNetV2 classifier is trained on Road Surface Classification Dataset (RSCD) dataset to predict a real-time pavement surface from 16 pavement surface classes (e.g., Wet Asphalt Severe, Dry Concrete Smooth), achieving 90% test accuracy and a macro F1-score of 88%. The image-based PSI (psi_image) for the predicted pavement surface class was effectively fused with the real-time weather-based PSI (psi_weather) using a Multi-Layer Perceptron (MLP) regressor. This regressor was trained on the Cartesian product of image and weather features, resulting in an impressive R² value of 0.999. The proposed multimodal solution combines Artificial Intelligence (AI)-based surface condition analysis and weather-based risk modeling to provide context-aware, real-time pedestrian safety, which enriches pavement research. This work offers a scalable method for proactive pavement maintenance and urban safety planning by changing the paradigm from vehicle-oriented indices to pedestrian-centric risk estimation.