Driver Pose Estimation by Hybrid Convolutional Network Architecture

In this paper, we introduce a novel and efficient method for articulated driver pose estimation in videos using a hybrid Convolutional Network Architecture, which incorporates both an optimized joint part detector and a joint-associated geometric constrained energy function. The structure of the part-detector algorithm, which uses the learned spatial context color and motion features that combined with the proposed single-person 2D joint association and Part Affinity Fields(PAF) method to establish a Gaussian model to improve the performance of the joint detection. At the same time, the energy function is used to constrain the detected driving pose to adapt to the more complicated interior environment. The method we proposed is significantly used in Advanced Driver Assistance Systems(ADAS) to help drivers avoid dangerous driving maneuvers. Finally, we illustrate our cascade structure accurate results in static image testing datasets and real-time on recorded driving scenes.