Driving confidence detection by machine learning classification and local outlier factor validation

The assessment of driver confidence is crucial in the Level 3 conditional automation driving (CAD) system defined by SAE J3016 and the advanced driver assistance system (ADAS). However, it is rather challenging due to reliance on subjective methods. This study presents an objective evaluation framework combining multi-sensor data acquisition with machine learning (ML) and statistical analysis. Data were collected from 38 drivers using an instrumented prototype electric vehicle (EV) equipped with a large suite of time-synchronized sensors including an Inertial Measurement Unit (IMU), a Force-Sensitive Resistor (FSR), a Galvanic Skin Response (GSR), a current sensor, a potentiometer, an encoder, and a Global Positioning System (GPS) receiver. In this study, three methodological approaches are tested and evaluated to assess driver confidence. Firstly, batch classification comparing drivers’ 38 × 80 feature vectors (mean, standard deviation, skewness, kurtosis) with respect to the driving safety expert is implemented. Euclidean distance is computed and ML algorithms are compared by evaluating the impact of sensor measurements. Secondly, stress detection using GSR biomarkers is tested and evaluated on an individual driver basis for anomaly detection purposes. GSR data is processed through moving window analysis. Finally, anomaly detection using the Local Outlier Factor (LOF) is validated by using position, speed, and heading measurements from a GPS receiver and a driving road scene captured by a camera. The methodology presents multi-sensor fusion for driver assessment, GSR-based stress detection, and LOF-based anomaly detection that has also been used to interpret ML predictions. Test results illustrate the effectiveness of the presented methodologies across modalities, 92.1% batch classification accuracy, 100% classification accuracy by using a current sensor, and 97.4% accuracy by using GSR. Driving anomaly is detected with a 98.6% accuracy on an individual driver basis. LOF identifies 100% of safety-critical instances that are also validated by the driving safety expert.