Research And Design Of Driver Fatigue Detection System

With the development and growth of the transportation industry and the development of the autonomous driving industry,the vehicles driving on the road every day will have many unknown risks.Therefore,researchers in the field of vehicle safety have ABS,ESP,electronic handbrake,driver fatigue detection systems,etc.Extensive research has been conducted on vehicle active safety equipment.Driver fatigue detection technology is a key part of the vehicle’s active safety equipment,which aims at protecting driver and pedestrian’s property and life by monitoring driver’s status and warning.Firstly,the principle,advantages and disadvantages of the key technology feature extraction algorithm and the state classification algorithm of driver fatigue detection are studied,and finally deep learning is selected as the driver fatigue detection system’s feature extraction and state classification method.Then,a feature extraction algorithm and a state classification algorithm are designed based on the deep learning method,in which the feature extraction algorithm is used to extract the image of the driver’s face region.The Algorithm classifies five states: drowsiness,yawning,smoking,talking on the phone,and normal driving.The vulnerability of the deep learning algorithm is analyzed,and a generative adversarial network algorithm is designed as an attack module.The fatigue confrontation samples generated by it and the collected real samples of the driver train the state classification network at the same time,so that the classification network has the ability to distinguish the confrontation samples from the real samples,and improve the defense ability against the fatigue counter samples.Secondly,a driver fatigue detection framework that cascades a convolutional neural network and a generated confrontation network is proposed,which mainly includes an information collection module,a feature extraction module,an attack module,a state classification module,and a judgment and early warning module.Information acquisition module: the real-time State video of the driver is collected by the infrared camera,and then the frame image representing the driver’s state is obtained from the video;State classification module: take the driver image as input,obtain the driver’s face range and key point coordinates,and then divide and extract the driver’s face according to the obtained face and key point position to obtain the driver’s ear,image of the area of the mouth and eyes;Attack module: use generative adversarial network technology to generate driver’s fatigue countermeasure samples through noise,that is,pseudo doze and pseudo yawn images,and mark them as non-fatigue state;State classification module: use real samples and fatigue confrontation samples to train together,so that the state classification module has the ability to resist pseudo-fatigue image attacks while distinguishing driver states.It can classify five driver states,including normal driving,drowsiness,yawning,making phone calls,smoking;Judgment and early warning module: take the state-classified image as input,according to PERCLOS algorithm and driver characteristics judgment within a period whether the driver is in fatigue state,and record and warn the fatigue state..Finally,the hardware solution of the fatigue detection system is studied,and the use of the Raspberry Pi 3B + and Intel neural computing stick as a hardware solution is proposed,which meets the needs of algorithm operation and has a high cost performance.Then the feasibility of the proposed cascaded convolutional neural network and the driver fatigue detection system for generating an adversarial network are verified,including the verification of the accuracy of the fatigue detection system and the attack using malicious images to verify the robustness of the system.Experiments show that in this paper,in terms of defending against fatigue against sample attacks,the state classification net work trained against samples has increased the detection rate of pseudo-fatigue images from 15% to 84%.The proposed area extraction algorithm has an extraction accuracy rate of more than 94% for the driver’s facial features.The fatigue detection system has achieved an accuracy rate of 92% in the detection of the driver’s fatigue state.In summary,the hardware scheme proposed in this paper is more accurate and can meet the requirements of the fatigue detection system and the operation system.