Research On Fatigue Driving State Detection Algorithm Based On MobieleNet-V3

From the data provided by the WTO,it can be found that there are countless traffic accidents every year and fatigue driving is one of the important reasons Among the 47 accidents mentioned in China’s traffic accident statistics report in 2005,7 were caused by fatigue driving.As we all know,fatigue driving is not only hurting yourself,but also hurting others.In order to avoid fatigue driving accidents as much as possible,fatigue driving state detection is extremely important.Therefore,the research on fatigue driving state detection has profound significance and pivotal value.In order to effectively carry out real-time detection during driving,this paper makes an in-depth research on fatigue driving state detection based on mobilenet-v3 lightweight network model so as to design an accurate and fast real-time method to recognize fatigue driving state.Accordingly,the main work indicated as below:(1)Aiming at the problem that the Squeeze and Excitation attention module in the MobileNet-V3 lightweight network model has too much calculation and parameter overhead and is not suitable for compact convolutional neural networks,the research choose an ultra-lightweight subspace attention mechanism module to replace.The ULSAM attention mechanism infers different attention maps for each feature map subspace to realize multi-scale and multi-frequency feature representation,which is conducive to fine-grained image classification.The accuracy can reach 97.53% and FPS can reach 24.5.Compared with SE attention module and convolutional block attention module,it improves by 2.5% and 13.4% respectively.(2)Although the attention module can show better performance,it only considers the current characteristics each time and does not share information with each other.In addition,this method does not exert the maximum effect of the attention mechanism.Therefore,this research adds a deep connected attention network module to connect the adjacent ULSAM lightweight attention modules,so as to make the feature information flow continuously in the channel block and enhance the attention model.The accuracy can be increased by 0.21%.(3)The improved MobileNet-V3 is used as the backbone network of YOLOv3 to detect the target face and determine five key points which including left and right eyes,nose,and mouth corners.Besides,the research use the nadam optimization algorithm,cyclic learning rate and biasloss loss function as optimized training conditions to pretrained MobileNet-V3 network model,and than recognize the state of the eyes and the mouth.Finally,the percentage of eye closure time and the frequency of yawning can be calculated by the percentage of eyelid closure over the pupil algorithm and judge whether the driver is in fatigue.The final accuracy can reach 94.41% and the average FPS can reach 27.73.This method meets the requirements of fast and accurate real-time detection of fatigue driving status.