Driver Fatigue Detection Based On Spiking Neural Network

Traffic accidents pose a great threat to people’s property and personal safety.According to the WHO,fatigue is responsible for 40% of traffic accidents.Therefore,it is important to detect fatigue and provide timely warning to drivers.Among the current methods for driver fatigue detection,convolutional neural networks dominate due to their powerful image recognition and classification capabilities,but the drawback of this method is that it takes up a lot of resources and consumes a lot of energy.As the latest generation of neural network,spiking neural network is in line with the neuronal mechanism of human brain,and it is lower than convolutional neural network in both computation and energy consumption,but the training process of spiking neural network is not easy.For this reason,in this thesis,the trained convolutional neural network weights are migrated to the spiking neural network method for fatigue detection,this method can combine the advantages of both,while solving the problem of difficult training of spiking neural network and implementing simulation on spiking neural network.The main work of this thesis is as follows:(1)Spiking neural network driver face detection method based on parameter migration transformation.For the problem of weak recognition features of spiking neural network after migration,the frequency coding method based on attention mechanism is proposed as a way to ensure the maximum spike issuance rate.The parameter migration transformation method is based on the YOLO model and migrates its parameters into the spiking neural network,proposes implementation methods for convolutional layers and normalization,and validates the effect of parameter migration as well as coding methods on the face dataset.The experimental results show that the accuracy of the migrated spiking neural network reaches 91.96% on the face detection dataset,which is similar to that of the original network,and the migrated network model consumes less resources and consumes less energy.(2)Construction of a deep spiking convolutional neural network-based eye and mouth image classification model.The model combines both the low-power spiking neural network and the deep convolutional neural network with good classification capability.The learning method used to train the model is a combination of two learning methods: unsupervised learning with spike time-dependent plasticity rules and reinforcement learning with reward modulated spike time-dependent plasticity rules.The experimental results show that the model achieves96.53% and 95.75% accuracy for eye and mouth state recognition on the eye and mouth datasets,respectively,with an accuracy improvement of 3.17% and 1.5% compared to the convolutional neural network under the same experime ntal conditions,respectively.(3)Driver fatigue determination by combining facial and head information.To address the situation of inaccurate detection due to single fatigue determination index,five fatigue feature indexes,namely PERCLOS,eye closure time,blinking frequency,yawning frequency and head nodding rate,are defined and combined to determine whether the driver is fatigued.Experiments show that the fatigue detection accuracy reaches 92.31% on the Yaw DD dataset,which is 2.01% less accurate than the YOLOv3 algorithm,but the energy consumption is only one-fifth of it.The low energy consumption of the spiking neural network is suitable for realtime driver fatigue detection and scenarios with limited computational resources.