Deep CNN, Body Pose, and Body-Object Interaction Features for Drivers’ Activity Monitoring

Automatic recognition and prediction of in-vehicle human activities has a significant impact on the next generation driver assistance intelligent autonomous vehicles. In this article, we present novel single image action algorithm inspired by perception that often focuses selectively parts images to acquire information at specific places which are distinct given task. Unlike existing approaches, argue activity is combination pose semantic contextual cues. detail, model considering configuration body joints, their interaction with objects being represented as pairwise relation capture structural information. Our body-pose body-object representation built be semantically rich meaningful, highly discriminative even though it coupled basic linear SVM classifier. We also propose Multi-stream Deep Fusion Network (MDFN) for combining high-level semantics CNN features. experimental results demonstrate proposed approach significantly improves drivers’ accuracy two exacting datasets.