Context-driven detection of distracted driving using images from in-car cameras

Distracted driving on roads has been a perennial problem. There is increasing consensus now among various stakeholders that this practice will not be eliminated with mere warnings from subject matter experts. Instead, these warnings must also be augmented with appropriate technological assistance to facilitate safer driving. There has been an urgent interest in the community to devise newer technologies for detecting when drivers are distracted on roads. In this paper, we design a computer vision techniques that process image data recorded from inside of cars to automatically infer when subjects are driving distracted. However, our innovation lies in adding context to the predictions. We do so, by first detecting and localizing a number of objects in cars that contribute to distracted driving (e.g., hands, smartphones, radio etc.) from images. Then, we process the relative locations of these objects using machine learning algorithms within an image to make predictions on distracted driving. We believe that our proposed context-driven approach is unique. We expect it to better facilitate correction of distracted driving, when real-time feedback to subjects come with appropriate contextual interpretation of the specific aspects that contributed to distracted driving. Performance evaluations of our techniques reveal a) mAP score of 63.90 for an IoU of 0.5 in object localization; and b) an overall accuracy of 94% in detecting instances of distracted driving based on object localization. Processing time incurred by our technique is around 200ms only. As such, we believe that our system is accurate, fast, practical and context-aware also.