Abnormal Driving Detection using Unsupervised Learning
Driver distraction is currently one of the major causes of traffic acci- dents. With the upcoming of in-vehicle information systems this problem is expected to become even bigger. To prevent this growth, methods to detect distraction and other abnormal driving behavior are needed. The aim of this study is to use non-intrusively gathered data to detect abnor- mal driving behavior in an unsupervised way, eliminating the need for a labeled dataset. Data from sensors that regard the vehicle and its sur- roundings were used. After an extensive data cleaning process 16 features were selected. Two state-of-the-art algorithms were used: a k-nearest neighbor algorithm (STORM) and an incremental local outlier factor al- gorithm (iLOF). The techniques were applied to data streams, making the learning and detection process online and real-time. In total 20 rides with overall 665 380 datapoints were investigated. The algorithms were not able to consistently find abnormal datapoints according to a qualitative analysis using video images as reference. In total only 7% of the outliers found by STORM and 8% of the outliers found by iLOF were considered to be qualitative anomalies. This study for the first time uses unsuper- vised outlier detection methods to detect distraction in automotive-based data streams. Important limitations found were that iLOF has a ten- dency to create too dense clusters when too homogeneous data is present, whereas STORM found mostly single isolated points. Main recommenda- tion for further study is to improve the quality of the data before further developing the application of these algorithms in this field.