This paper presents a comprehensive approach to accident detection using supervised and unsupervised learning methods. Our methodology encompasses data gathering, preprocessing, and dimensionality reduction, applied to a highly imbalanced dataset with 99.99% of instances being non-crash and 0.01% of instances being crash. We evaluate various traditional machine learning algorithms, which achieve high accuracy but fail to provide satisfactory precision, a critical metric for our application. To address this, we propose an unsupervised learning method—novelty detection—that effectively identifies crash instances without relying on labels, accurately predicting the crash events based on the provided “Crash Time.” Our results demonstrate the limitations of traditional supervised methods in this context and emphasize the potential of unsupervised learning for improved accident detection.