Driving State Detection By Human Posture Estimation And Gaze Attention

As the most common means of transport in human life,the automobile always maintains its strong and rigid demand.While it brings convenience to people's travel,it also leads to a year-on-year increase in the incidence of traffic accidents.Among all the causes of traffic accidents,bad driving habits and poor driving conditions are the most important factors.If we can predict in advance to determine the driver's behavior,perception of the driver's poor state analysis,and timely warning,so that the vehicle automatically slowed down and lit warning lights,can effectively reduce the incidence of traffic accidents.So,it is great significance to study reliable and effective driving condition detection technology to prevent traffic accidents.In our paper,aiming at the problem of driving condition detection,various methods in related fields at home and abroad are studied,and its advantages and disadvantages are summarized.Proposed from the human body posture estimation and attention from two aspects.The main contents include the following sixth sections:1.Based on monocular camera body's three-dimensional posture estimation.By using the monocular camera to get the image,extracting and combining the features of the human body on the two-dimensional image to form the outline of the human body,the similarity is compared with the projection of the pre-established three-dimensional human body model.By making the 3D model track the image estimate the whole body and upper body posture.2.Optimize the effect of three-dimensional attitude estimation.In view of the particularity of the environment in the vehicle,four issues of in-car squint angle,influence of drivers' clothing on the algorithm,staggered hands when using the steering wheel and imaging of the infrared camera at night are considered.An improved method of correcting the initial state of the model,setting the threshold classification,defining the error correction function and restoring the normal image by using the production confrontationnetwork are proposed,and the effect of the three-dimensional pose estimation in the cab is optimized.3.Sight tracking.Based on Real Sense,a three-dimensional eyeball model is established and the whole process of line-of-sight tracking is divided into two parts: static and dynamic.For dynamic line-of-sight tracking,a three-dimensional model of the head is established.For static line-of-sight tracking,the line-of-sight area is divided into eight directions,and the line-of-sight area is determined by the difference angle between the center point of the eyeball and the center point of the eye.4 face gesture estimation.With the help of Real Sense,78 positions of three-dimensional feature points on the face were obtained.Ten different face gestures were created by the combination of three-dimensional coordinates and attitude angles of multiple feature points.Corresponding cost functions were designed and ten face Behavior recognition.5.Driver behavior recognition.In this paper,the driving behavior is divided into two types: dominant and recessive.First,the dominant behavior is identified by the algorithm in the text.For the recessive behavior,the dominant behavior is combined with multiple explicit behaviors to determine the relevant judgments Process and cost functions.6.Driver status judgment.In this paper,the driving status is defined as three levels first,and then based on different driving behaviors and corresponding durations or occurrences,the driver's status is finally determined synthetically.