Real-time Machine learning-based approach for pothole detection

Potholes are symptoms of a poorly maintained road, pointing to an underlying structural issue. A vehicle's impact with a pothole not only makes for an uncomfortable journey, but it can also cause damage to the vehicle's wheels, tyres and suspension system resulting in high repair bills. This study presents a comparative study of machine learning models for pothole detection. The data was collected from multiple android devices/routes/cars and pre-processed using a 2-second non-overlapping moving window to extract relevant statistical features for training a binary classifier. The Test dataset was isolated entirely from the Training and Validation datasets, and a stratified K-fold cross-validation was applied to the Training dataset. The Random Forest Tree and KNN showed the best performance on the Test dataset with a similar accuracy of 0.8889. The model performance increased when random search hyperparameter tuning was applied to optimise the Random Forest Tree model's hyperparameters. The Random Forest Tree model's performance after hyperparameter tuning is 0.9444, 1.0000, 0.8889 and 0.9412 for accuracy, precision, recall, and F-score, respectively.