Stay As You Were!: Unsupervised Driver Behavior Profiling through Discovering Normality on Smartphone Sensor Measurements

Abstract

Driver behavior profiling is a significant technology in intelligent transportation as it provides contextual knowledge regarding the driver's aggressiveness. The prior studies analyzed the data's temporal characteristics and established classifiers between the normal and aggressive driver behavior in a supervised manner. However, there exist limits that the practitioner should acquire a labeled dataset, and the model could not identify unseen driver behaviors a priori. To hedge the aforementioned limits, our study proposes a novel driver behavior profiling approach under the normality discovery paradigm, which is unsupervised learning. First, we presented practical feature engineering steps to transform the smartphone IMU's raw sensor measurements to the sequence of driving data. Second, we established an unsupervised driver profiling approach that necessitates the driving data of normal driver behavior only for the model training. Third, we figured out each aggressive driver behavior has a different sequence length to represent its unique patterns. Lastly, we compared our approach's performance with a supervised approach and resulted in our unsupervised model achieved similar performance in identifying aggressive right turn, left turn, and left lane change, but required further improvements in recognizing an aggressive left lane change, aggressive braking, and aggressive acceleration.