Overview of the Project

Recent studies on naturalistic driving data suggest that driver state (attentive, distracted, stressed/angry, sleepy, fatigued, limited mobility) can significantly contribute to the risk of crash. A recent study on driver crash factors and prevalence using driving data found that driver-related factor are present in around 90% of crashes. Current techniques for data reduction require manual identification of activities such as distraction, sleepiness etc. and hence can only be applied to a limited study sample.

This grossly limits the ability to understand the role of driver state variables on crash risk. There is a need to automate the driver state estimation to fully utilize the available naturalistic study data.

To address the above limitation, we propose to design Deep Insight model, which is a robust deep net based driver state estimation system using camera and other in-vehicle sensors. Figure 1 provides an overview of the proposed Deep Insight solution. The team will design and develop a high performance computing driven Deep Insight platform to support end to end model development and deployment.  Driver state is defined by a combination of physiology (activity, and sleep), awareness of internal state, cognition, medical state (restricted mobility).

Innovative Ideas

Deep InSight’s innovative model will include the following features:

• Combining multiple naturalistic video datasets.
• Deep Insight Artificial Intelligence (AI) platform.
• Enhancing Image Quality of naturalistic driving recordings.
• Face detection at extreme angles.
• Advanced driver state estimation using long short term memory (LSTM) networks.
• Finding the factor of improvement with additional sensors.