Inaccurate Prediction Is Not Always Bad: Open-World Driver Recognition via Error Analysis

Driver identification is of fundamental importance in many vehicle-related applications, such as fleet monitoring and anti-theft system. The vast majority of existing methods work under the closed-world assumption, which may be unrealistic in practice. In this paper, we consider a more practical but challenging scenario, i.e., open-world driver recognition, and propose a systematic method dubbed DRIVERPRINT. To recognize the driver of interest, DRIVERPRINT takes advantage of the behavioral predictability of the driver himself, thereby no need to collect data from other drivers for model training. Specifically, DRIVERPRINT predicts the behavior-related traveling speed with a driver-specific predictor, compares the prediction error with a pre-trained error model and finally recognizes drivers via error analysis. Besides open-world setting, our method is also compatible with closed-world driver classification. Real-world experiments demonstrate our method achieves reasonable accuracy. The average F1-score for open-world driver recognition is up to 0.91, while that for closed-world driver classification is up to 0.973.