Research On Fatigue Driving Detection Algorithm Based On Computer Vision And Deep Learning

Fatigue driving is an important cause of traffic accidents,so implementing effective fatigue driving detection is particularly important.Currently,fatigue driving detection algorithms suffer from poor comfort,susceptibility to external factors(such as lighting conditions,mask wearing,and sunglasses),low accuracy,and poor real-time performance.Most existing fatigue detection algorithms are based on a single network,making it difficult to accurately detect all facial features of the driver when they are in a fatigued state,as different drivers may exhibit different fatigue states.To address these issues,I have designed a fatigue driving detection system based on computer vision and deep learning.The system includes a face detection module,head pose estimation module,eye and mouth state judgment module,and multi-factor fusion decision-making module.First,the face detection module detects the driver’s face and five feature points,and then extracts the driver’s eye and mouth areas based on the feature points.The detected driver’s face image and extracted eye and mouth areas are then passed separately into the head pose estimation module and the eye and mouth state judgment module to calculate the driver’s head pose angle and determine their eye and mouth state.Finally,the multi-factor fusion decision-making module is used to comprehensively determine the driver’s fatigue state.The main work and innovations of this article are as follows:1.In the face detection module,this thesis designed a lightweight face detection network.This network reduces computational complexity,improves inference speed,and ensures accuracy,enabling real-time and accurate face detection and alignment during driving.The network consists of three parts: feature extraction,feature fusion,and multi-task loss function.I have designed a new loss function that,in addition to performing face bounding box regression,can also regress the five facial feature points,which helps with face detection.The new loss function can make the model converge faster,ensure more stable face bounding box regression,and improve the accuracy of the network.2.In the head pose estimation module,this thesis designed a network for calculating the driver’s head pose angle.This network transforms the head pose regression problem into a classification problem and adopts a coarse-to-fine classification strategy.Compared with the mainstream method of using CNN to extract2 D facial landmarks,this network has lower computational complexity,higher accuracy,and faster detection speed.3.In the eye-mouth state judgment module,this thesis used a Transformerthat has a better ability to capture global information instead of CNN.The Transformerutilizes its self-attention mechanism to better learn the global information of the image.To fully utilize the contextual information of the image,we improved the self-attention mechanism in the Transformer.First,we encoded the key with contextual information and then generated the attention matrix through convolution.4.In the driver fatigue detection system,multiple facial features of the driver,such as changes in blink frequency,increased yawning,longer eye-closing time,and improper head posture,exhibit obvious changes when the driver is fatigued.Therefore,based on three facial features,namely,the eyes,mouth,and head posture,this thesis designs the MF-Algorithm to comprehensively determine whether the driver is fatigued.The system uses the output of the head posture estimation module and the eye-mouth state judgment module to determine the driver’s fatigue state based on multiple factors,including blink frequency,ratio of eye-closing time,duration of a single eye-closing event,yawning time ratio,and head posture.5.Using multiple modules to detect the facial features of drivers and integrating the information from these modules for comprehensive assessment can effectively reduce the interference of external factors.The interaction among these modules can help address the issue of occluded eyes or mouth caused by wearing masks or sunglasses while driving.The entire system is based on computer vision and image processing techniques to obtain facial features of the driver.The system only requires an RGB camera to be placed in front of the driver,ensuring their comfort.By using multiple networks in combination,the impact of external factors can be reduced,and the use of multiple facial features of the driver can improve the accuracy of the system.The effectiveness of each module has been demonstrated on the WIDER FACE and BIWI public datasets.Additionally,the entire system was tested on a self-built dataset,achieving an accuracy of 97.8%.