Master's Thesis Defence | Brian Mao, Road Surface Estimation Using Machine Learning

MS Teams: Please email amgrad@uwaterloo.ca for meeting link

Candidate

Brian Mao | Applied Mathematics, University of Waterloo

Title

Road Surface Estimation Using Machine Learning

 Abstract

Vehicle motion control systems are present on commercial vehicles to improve safety and driving comfort. Many of these control systems could be further improved given accurate online information about the road condition to accommodate for driving under poor weather conditions such as icy roads or heavy rain. However, sensors for direct friction measurement are not present on commercial vehicles due to production costs. Hence, it is beneficial to incorporate an online estimation scheme for road surface classification.

This thesis focuses on investigating two fundamentally different machine learning-based methods for road surface classification. The first is an artificial neural network that provides a global function approximation of the underlying dynamics. In particular, Long Short-Term Memory (LSTM) units are used to capture temporal relationships within the training data and to mitigate the vanishing gradient problem. The second is an instance-based learning method referred to as Nadaraya-Watson Kernel Regression, where local function approximations are generated around the input data. 

Results indicated that both machine learning-based methods were able to classify road conditions to a reasonable degree of accuracy after tuning associated hyperparameters. However, each method has different benefits and drawbacks. The LSTM network model was capable of making accurate predictions on inputs drastically different from data points within the training data set, was generally more accurate on predictions associated with new driving maneuvers, required less storage for implementation, and had relatively short prediction times. Conversely, the Nadaraya-Watson Kernel Regression model was generally more accurate at making predictions on inputs that were very similar to data points within the training data set, did not require any training time to incorporate newly collected data into the model, and generated predictions that were more easily explainable.