Driver Activity Recognition by Fusing Multi-object and Key Points Detection

Driver distraction recognition plays a fundamental role in road safety. In this paper, we present a modular architecture based on the fusion of key points and object detection for predicting driver's actions. From multi-camera infrared recordings, we will temporarily detect among a variety of actions that lead to distractions. Our system detects objects of interest and extracts key points from the driver. They are merged by generating features that relate them and processed with a ML-based classification algorithm. Finally, filters are applied to reduce bounces and add temporal context to the detections. Our proposal has been validated on two state-of-the-art datasets for driving distractions. Through several experiments we show that fusion substantially improves related action inference and improves domain adaptation. In addition, our framework is lightweight, explainable and has a low latency as it performs frame-by-frame inference. The modularity of the network allows us to upgrade parts independently or eliminate a camera without having to modify the entire network.