Study On Fatigue Detection System Based On Multi-task Coordination Deep Learning

Fatigue driving is one of important reasons which lead to traffic accidents.How to effectively detect the states of driving fatigue and give warnings,has become a hotspot of scientists from all over the world.Compared with the fatigue detection methods using physiological parameters,the non-contact methods based on machine vision has attracted more and more attention.However,how to effectively improve the accuracy and stability of fatigue detection is still valuable for us to further exploration under the complex environment of the illumination variations,partial collusions and vibration.In this paper,some research work is carried out based on deep learning and multi-task coordination.First,a deep convolution neural network(DCNN)model is designed to detect facial landmark,and an auxiliary training method is used to improve the feature learning ability of the model.Then synergistic optimization of both facial landmark localization and head pose estimation tasks is studied,and the geometric constraint approach is employed to pre-train the DCNN.Finally,a decision-making algorithm of multi-feature fusion is proposed to estimate fatigue states.The effectiveness of the proposed methods is verified by experiments in public datasets(300W)and actual test data.The main work and research results are listed as follows:(1)Considering that cascade DCNN has too many network layers,model parameters and inadequate training of underlying parameters,a detection method of facial landmark is proposed based on assistant training DCNN.A single network is first used to estimate facial landmark,which has lower complexity than cascade network and improves the running efficiency of the algorithm,and then the auxiliary training technique is applied to enhance the feature learning ability of the model.Compared with traditional approaches in experimental results,the proposed method has achieved better precision and real-time performance.(2)In order to address the problem that the feature learning ability of single-class data source is limited in the single-task deep network,and DCNN network initialization is difficult,a learning method is proposed based on multi-task coordination and the initialization of geometric constraints.First,due to the correlation between facial landmark localization and head pose estimation,these tasks are optimized jointly to simultaneously estimate the coordinates of facial landmark and the angles of head pose.Then,the geometric constraint algorithm is employed to pre-train the DCNN.The proposed approach not only extracts the invariant features of pose variation,but also speeds up the training time of the network.The experimental results show that the proposed method can effectively improve the detection accuracy under the condition of large pose change.(3)For the problem that existing fatigue detection systems mostly adopt single characteristics of fatigue state which lacks of comprehensiveness and has weak adaptive capacity in complex environment,a decision-making method of multi-feature fusion is proposed.Multi-task coordination simultaneously estimates the coordinates of facial landmark and the angles of head pose.Combing the two values,multiple features which are eye state,mouth shape and head pose can be used to assess driver fatigue,in order to improve the estimation accuracy.In the actual driving environment,test data is collected and analyzed,and experimental results show the feasibility of the proposed algorithms.