Driving Posture Detection using Depth Sensor

Driver Monitoring Systems (DMS) estimate drivers’ states by exploiting sensors, such as image sensors and biosensors. An understanding of drivers’ states enables advanced driving assistances. For example, existing DMS recognizes fatigue or drowsiness and alert the drivers. If DMS are widely used, the number of traffic accidents due to human errors, which account for the majority of traffic accidents, will be decreased. Changes in the modes of traffic require the estimation of whether drivers are able to control their vehicles in some situations.
For example, in automated driving, there is a need for estimating the driver’s readiness. According to SAE International, automated driving is divided into five levels of automation. At each level, the transition of driving authority from an automated vehicle to its human driver will occur. Level 3 vehicles, which are semi-automated, require drivers to take over the controls in cases of system failure. Even at the fully automated levels, levels 4 and 5, there will be transitions when drivers want to drive, for example, for excitement. In order to safely transfer control, the driver must be in a proper state to drive the vehicle. Therefore, estimating the driver’s readiness is essential in automated driving.
Another example is manual driving, in which the estimation of the driver’s readiness is also needed. Recently, the number of traffic accidents due to drivers’ health states has been increasing. When drivers lose consciousness because of health issues, they lose control of the vehicle, which result to a serious accident. For example, in Japan, the death rate is six times higher for these cases than for other accidents. Not just in automated driving but also in manual driving, the estimation of the driver’s readiness is needed.
To estimate the driver’s readiness, certain studies provide helpful advice; drivers are able to take evasive actions in optimum driving posture and driver posture analysis is useful for accident prevention.
We have proposed a DMS to estimate the driver’s readiness by understanding driving posture. We use the idea of anomaly detection, which creates normal models from datasets and treats outliers from the normal models as abnormal. In our DMS, when the driver’s posture is not similar to that of normal driving, the system decides that the present driving is abnormal, or that the driver does not have control ability.
We substantially have improved the accuracy of this DMS by introducing the semantics of the relative relationships of the joint positions. Based on the notion that there are different relative relationships of the joint positions between normal and abnormal driving, we re-define the distance metric of driving postures. Our system is implemented on an embedded GPGPU platform, so we achieve real-time performance by parallel computing on GPGPU. Our novel DMS was evaluated in a real vehicle with 10 participants. The subjects drove the vehicle and we checked our DMS performance in real time. The evaluation results show that the average false detection rate decreases to 0.01 %, so we confirmed that this novel DMS has potential performance for practical use.