A Research On Drowsiness Detection And Emotion Recognition Of Drivers

The widespread use of cars,not only bring convenience to our life,but also make driving safety is widely concerned.Fatigue driving and bad emotions are both important causes of traffic accidents,so it is of great significance to give early warning to the driver’s abnormal state.Aiming at the main difficulties of these two problems,this paper studies the effective drowsiness detection and emotion recognition algorithm,in order to accurately obtain the fatigue and emotion state of the driver,to provide input for the safety early warning system.The innovative work of this paper includes:A Hierarchical Hidden Markov Model on binocular pair for driver drowsiness detection is proposed in this paper.Firstly,a target detection network is used as the eye detector to obtain the open and closed state of human eyes as the drowsiness-related feature.In this process,the position constraint is imposed by adding an outer bounding box to couple the separate area of two eyes;Secondly,a Hierarchical Hidden Markov Model is constructed by the eye state sequence with different sampling frequencies to extract fatigue information at specific time levels,which can represent the fatigue process of human eyes more accurately.Compared with several classical drowsiness detection algorithms,our method in this paper shows better performance,which achieves 92.98% accuracy on the NTHU-DDD dataset.This paper proposes an emotion recognition algorithm based on spatially transformed Bayesian Convolutional Neural Network.Firstly,the spatial transformation network is used to calibrate the pose of the original facial expression image,which is helpful to extract the efficient expression features;Secondly,facial expressions have great uncertainty,so a Bayesian Convolutional Neural Network is constructed as a classification model for facial expression recognition.The model is introduced with uncertainty by making the network parameters obey some probability distribution,which can enhance the expressiveness of the model.Through comparative experiments,it is proved that the accuracy of the model in this paper surpasses the current several representative expression recognition models.The proposed emotion recognition algorithm achieves 99.1%,82.3% and 83.5% accuracy on KMU-FED dataset,RAF-DB dataset and FERplus dataset respectively,without using the pre-trained model.