Optimization of extreme position landmarks identification algorithm for driver fatigue state

Abstract

Aiming at the problem of poor fitting effect of key points of the driver's mouth and eyes during driver fatigue monitoring, a convolutional neural network PFLD-rep based on improved PFLD was proposed. First, the backbone network is structurally re-parameterized, which can not only retain multi-scale parameters but also achieve the calculation speed of a single branch; then introduce the idea of heat map regression to the auxiliary network to improve the spatial expression ability of the network in the mouth and improve fitting. Effect; Add the opening angle of the eyes and mouth to the loss function, and replace the original L2_loss with Wing_loss loss to increase the network's attention to mouth deformation; later, a high-precision comprehensive fatigue judgment criterion will be determined through experiments. The final experimental results show that the IPN value of the PFLD-rep algorithm in the 300W data set reaches 3.68. The algorithm also has a high fitting effect in faces with large deformations, and can be implemented when vehicle-mounted hardware equipment requirements are low. Higher precision monitoring effect.