While vehicle automation increases in the near future human drivers will still be responsible for monitoring of the driving environment or as a fallback for critical situations. Thus, systems for autonomy levels 2 to 3 require proper knowledge about the driver's state including, for example, the seating position, activity, or hands-on-steering-wheel status. We introduce an interior sensing system to classify the state of the driver inside a vehicle based on images of a single time-of-flight camera. We present an efficient and robust approach based on a hierarchical label structure and a cost function which focuses only on relevant labels inside this hierarchical label structure. Our approach reduces the number of output neurons while creating a fallback option inside the classifier simultaneously. This leads to improved classification results compared to common approaches. In our experiments we demonstrate a solution for combined assessment of seat occupancy, driving position, hands on steering wheel, and driver-object interaction.
Driver State Monitoring with Hierarchical Classification
2018-11-01
2436692 byte
Conference paper
Electronic Resource
English