A framework for context-aware driver status assessment systems
The automotive industry is actively supporting research and innovation to meet manufacturers' requirements related to safety issues, performance and environment. The Green ITS project is among the efforts in that regard.
Safety is a major customer and manufacturer concern. Therefore, much effort have been directed to developing cutting-edge technologies able to assess driver status in term of alertness and suitability. In that regard, we aim to create with this thesis a framework for a context-aware driver status assessment system. Context-aware means that the machine uses background information about the driver and environmental conditions to better ascertain and understand driver status. The system also relies on multiple sensors, mainly video and audio. Using context and multi-sensor data, we need to perform multi-modal analysis and data fusion in order to infer as much knowledge as possible about the driver. Last, the project is to be continued by other students, so the system should be modular and well-documented.
With this in mind, a driving simulator integrating multiple sensors was built. This simulator is a starting point for experimentation related to driver status assessment, and a prototype of software for real-time driver status assessment is integrated to the platform.
To make the system context-aware, we designed a driver identification module based on audio-visual data fusion. Thus, at the beginning of driving sessions, the users are identified and background knowledge about them is loaded to better understand and analyze their behavior.
A driver status assessment system was then constructed based on two different modules. The first one is for driver fatigue detection, based on an infrared camera. Fatigue is inferred via percentage of eye closure, which is the best indicator of fatigue for vision systems. The second one is a driver distraction recognition system, based on a Kinect sensor. Using body, head, and facial expressions, a fusion strategy is employed to deduce the type of distraction a driver is subject to. Of course, fatigue and distraction are only a fraction of all possible drivers' states, but these two aspects have been studied here primarily because of their dramatic impact on traffic safety.
Through experimental results, we show that our system is efficient for driver identification and driver inattention detection tasks. Nevertheless, it is also very modular and could be further complemented by driver status analysis, context or additional sensor acquisition.