Turn-Intent Analysis Using Body Pose for Intelligent Driver Assistance

Human-centric, pervasive computing environments, with integrated sensing, processing, networking, and displays, provide an appropriate framework for developing effective driver-assistance systems. Also essential when developing such systems are systematic efforts to understand and characterize driver behavior. In an attempt to make such a predictive turn-assistance safety system a reality, we equipped an experimental vehicle with cameras and sensors to capture the vehicle dynamics, view of the road ahead, and driver's body pose. We investigated how and to what extent we could use body-pose information to detect and predict driver activities. We analyzed the detection performance of a two-class pattern classifier using receiver-operator-characteristic curves, which describe the classifier's ability to suppress missed detections and false alarms. The curves provide a ratio indicating the system's attainable proactivity (ability to foresee a user's needs) versus its transparency (ability to avoid user annoyance). Our goal is to eventually develop vision-based body-pose-recovery and behavior-recognition algorithms for driver-assistance systems