- 32 views
- 56 downloads
Harnessing Visual Analytics for Enhanced Winter Road Safety: A Framework for Continuous Road Surface Friction Estimation
-
- Author / Creator
- Xie, Qian
-
Winter Road Surface Condition (RSC) monitoring currently relies on qualitative descriptors such as "bare lane," "partially snow-covered," and "fully snow-covered. These descriptors present two inherent problems—their subjective nature, which gives rise to measurement inconsistencies, and the difficulty for road users to discern between safe and unsafe roads due to the broad spectrum of conditions intermediate RSC classes cover. Friction-based measurements, ranging from 0 to 1, alleviate these issues as they provide a much more objective measure. Despite this, the large-scale implementation of such measurements has long been impeded by high collection costs. Even with the incorporation of friction values, the existing RSC monitoring system is often plagued by limited spatial coverage and infrequent updates. This can result in potential discrepancies between reported and actual conditions, thus compromising the reliability and usability of the system.
To overcome these challenges, this thesis capitalizes on the advancements in computing capabilities and sophisticated machine learning techniques. The primary focus is to repurpose the information-rich image datasets and to enable continuous monitoring of dynamic winter RSCs. This strategy ensures a consistent flow of reliable and timely information for maintenance personnel and road users. At the core of this novel approach is a comprehensive framework designed to convert winter road surface images into friction values. This framework synergistically integrates three essential components—an image-based friction model, a friction interpolator, and a binary collision model. Together, they form a robust system aimed at addressing the prevalent issues in winter RSC monitoring for improved winter road safety.
For the development of the image-based friction prediction model, 128 friction measurements and their corresponding road surface images were collected over select roads in the city of Edmonton. Since the images themselves could not serve as direct inputs, feature extraction techniques were used to summarize the information found within the images. These techniques encompass RSC labeling into four classes (bare lane, one-track bare, two-track bare, and fully snow-covered), image thresholding, Local Binary Pattern (LBP), and Gray Level Co-occurrence Matrix (GLCM). With these extracted features, three tree-based algorithms; namely, decision tree, Random Forest (RF), and Gradient Boosting (GB), were used to model the relationship between the friction values and the extracted feature. All three tree-based models displayed robust performance based on standard statistical measures such as RMSE and RMSPE.
In an effort to improve the spatial coverage of RSC monitoring, the friction interpolator was subsequently constructed. Continuous friction values were generated by feeding collected road surface images into the friction model. Ten datasets were then created by varying the distance between available observation points from 100 m to 1000 m per observation. A total of six interpolators were evaluated on these datasets, including Ordinary Kriging (OK), Regression Kriging (RK), Random Forest (RF), RFOK, Random Forest Spatial Interpolator (RFSI), and RFSIOK. Among these interpolators, OK yielded the highest accuracy and displayed the least sensitivity across nearly all observation distances. An additional case study was conducted using the City of Edmonton’s traffic camera locations as a real-world scenario to demonstrate the applicability of the proposed model. This case study assesses the potential for using traffic camera data to interpolate friction measurements. Identical to the sensitivity analysis, OK was found to be the top-performing interpolator, achieving commendable accuracy with only five friction measurements as input.
To underscore the potential of the proposed framework in enhancing road safety, a binary collision likelihood model was developed by utilizing continuous friction measurements. After examining segment lengths ranging from 500 m to 20 km, the 6.5 km model was determined to offer the optimal balance between interpretability and accuracy. The developed model demonstrated robust performance leveraging Annual Average Daily Traffic (AADT) and friction as predictors. Furthermore, an additional analysis was conducted assuming the availability of Connected Vehicles (CV) and appropriate sensing technologies. This analysis took a more pragmatic approach to assess potential safety benefits by evaluating the predictive capabilities of the generated continuous friction measurements in relation to collision events. The findings offered a tangible demonstration of the potential safety improvements that could be realized using the proposed framework.
-
- Graduation date
- Fall 2023
-
- Type of Item
- Thesis
-
- Degree
- Master of Science
-
- License
- This thesis is made available by the University of Alberta Libraries with permission of the copyright owner solely for non-commercial purposes. This thesis, or any portion thereof, may not otherwise be copied or reproduced without the written consent of the copyright owner, except to the extent permitted by Canadian copyright law.