Learning When Agents Can Talk to Drivers using the INAGT Dataset and Multisensor Fusion
summary
Uses the INAGT dataset of 1,915 synchronized samples (in-cabin and road video, CAN telemetry, physiology), resampled to 10 Hz and segmented into 6 s clips; drivers answer Yes 78% / No 22%. Trains a multi-channel network PazNet (I3D Inception V1 for road, CNNs for sensors and landmarks extracted with OpenPose), applies SMOTE, and scores 0.874 F1 for driver preference and 0.891 F1 for third-party safety. Conservative tuning drops false positives to zero while still yielding ~24 interaction opportunities per 50-minute drive.
AI-written summary · about
chapters (11)
- 0:00Intro
- 0:23Proactive speech interfaces for cars
- 0:49Modeling Approach
- 1:32Sensor Signals
- 1:43Third-party safety ratings
- 2:08Preprocessing & Synchronization
- 2:33Multisensor network architecture
- 3:17Comparing Modeling Techniques
- 4:07Reweighting classes during training
- 4:38Modeling Safety
- 5:34Contributions
related talks
19mWhat Can Be Predicted from Six Seconds of Driver Glances?
08mHuman–AI Collab for Reliable Automotive Control Systems
36mSession on 'Learning to learn'
19mAI Trends in Intelligent Transport Systems
41mAI applications in autonomous driving
40mTeaching a Car to Drive Itself by Imitation and Imagination
31mDeep learning for self-driving vehicle technology
05mInterview