Mingming Zhao1,2,3,4, Georges Beurier1,2,3, Hongyan Wang4, Xuguang Wang1,2,3. 1. Université de Lyon, Lyon, France. 2. Université Claude Bernard Lyon 1, Villeurbanne, France. 3. Université Gustave Eiffel, UMR_T9406, LBMC Laboratoire de Biomécanique et mécanique des chocs, Lyon, France. 4. School of Automotive Studies, Tongji University, Shanghai, China.
Abstract
OBJECTIVE: Driver posture monitoring is useful for evaluating the readiness to take over from highly automated driving systems as well as for designing intelligent restraint systems to reduce injury. The aim of this study was to develop a real-time and robust driver posture monitoring system using pressure measurement. METHODS: Driver motion and pressure measurement were collected from 23 differently sized participants performing 42 driving and non-driving activities. Nine typical driver postures were identified by analyzing trunk and feet positions in 3 D space for classification. One deep learning classifier and two Random Forest classifiers were trained respectively on pressure distribution, absolute and relative pressure features extracted from pressure measurement. Leave-One-Out cross-validation was performed to evaluate the performance of the classifiers. RESULTS: Without considering feet positions, all the classifiers could provide reliable recognition of the normal trunk position for standard driving with an accuracy around 98%. With help of a reference sitting position, the best performance was achieved by Random Forest classifier trained on the relative pressure features with an average classification accuracy of 80.5% across 9 typical postures and 23 drivers. The main errors were related to the recognition of feet positions when applying braking and relaxing both feet on the floor. CONCLUSIONS: Pressure measurement could be a good alternative or complementary to camera based driver postural monitoring system. Results show that all classifiers proposed in the work could predict the trunk position for standard driving. With help of an initial posture, Random Forest classifier with relative pressure features could classify trunk positions with high accuracy. However, further effort is needed to improve the accuracy of feet position prediction especially by adding more foot related task data.
OBJECTIVE: Driver posture monitoring is useful for evaluating the readiness to take over from highly automated driving systems as well as for designing intelligent restraint systems to reduce injury. The aim of this study was to develop a real-time and robust driver posture monitoring system using pressure measurement. METHODS: Driver motion and pressure measurement were collected from 23 differently sized participants performing 42 driving and non-driving activities. Nine typical driver postures were identified by analyzing trunk and feet positions in 3 D space for classification. One deep learning classifier and two Random Forest classifiers were trained respectively on pressure distribution, absolute and relative pressure features extracted from pressure measurement. Leave-One-Out cross-validation was performed to evaluate the performance of the classifiers. RESULTS: Without considering feet positions, all the classifiers could provide reliable recognition of the normal trunk position for standard driving with an accuracy around 98%. With help of a reference sitting position, the best performance was achieved by Random Forest classifier trained on the relative pressure features with an average classification accuracy of 80.5% across 9 typical postures and 23 drivers. The main errors were related to the recognition of feet positions when applying braking and relaxing both feet on the floor. CONCLUSIONS: Pressure measurement could be a good alternative or complementary to camera based driver postural monitoring system. Results show that all classifiers proposed in the work could predict the trunk position for standard driving. With help of an initial posture, Random Forest classifier with relative pressure features could classify trunk positions with high accuracy. However, further effort is needed to improve the accuracy of feet position prediction especially by adding more foot related task data.