In-vehicle Driver Posture Recognition And Prediction Research

With the rapid development of the automobile industry in our country,traffic accidents are becoming more frequent.Although the active and passive safety system can prevent vehicle collisions to a certain extent and slow down the damage suffered by drivers,the number of drivers killed in accidents is still very large every year,and driver safety will always be the focus of research and development in the auto industry.Based on the current situation of the development of intelligent driving assistance system,not only for the traffic environment and the vehicle itself to provide real-time monitoring,but also more attention to the driver’s own driving posture and driving behavior,in order to achieve the current driving posture of the driver,and even the future driving posture can provide reasonable protection purposes.The main research content of this paper is as follows.(1)Driver upper body posture recognition.Referring to the automobile passive safety occupant injury guidelines and determining the human keypoint skeleton model,this paper uses the Kinect-Openpose algorithm as the basis to determine the human pose recognition process,and the OKO(Optimized Kinect-Openpose)algorithm of this paper is proposed after optimizing the problems such as low accuracy of Kinect and slow speed of Openpose.(2)Driver lower body pose regression.The driving simulator and CMM are used to collect the volunteer body size information,height and weight information,and seat stand parameters,and Pearson correlation analysis is used to find the parameters that become strongly correlated with the lower body,and the stepwise regression method is used to establish the lower body pose regression equation after inputting to SPSS software.For the upper body parameters may not be identified due to partial overlap of key points,this paper further distinguishes different occlusion cases to correct the regression equation to achieve the purpose of improving accuracy and applicability.(3)Driver whole-body accuracy verification.The accuracy of the driver’s wholebody pose model is verified after unifying the coordinate system of CMM and Kinect,including the accuracy of keypoints and joint segments;the accuracy of key points of this paper is compared with that of Kinect and Kinect-Openpose algorithm;and the accuracy of two skeleton models of Openpose is also verified;and the accuracy of the regression equation established under various occlusions of the lower body is verified.The accuracy of the regression equation is verified.(4)Driver’s keypoint motion trajectory prediction.The most realistic keypoint trajectories of the driver under the emergency braking and left and right steering conditions were collected in the real vehicle experiment,and the keypoints to be predicted under the three conditions were determined by analyzing the occlusion and motion amplitude of each part of the human body during the motion process in turn.The LSTM key point trajectory prediction network is established,and the optimal combination of hyperparameters is selected for each working condition using a multilayer grid search hyperparameter optimization method to upgrade keypoint pose recognition to keypoint pose prediction.The proposed driver pose recognition and prediction method can be accomplished by a single Kinect sensor,which is low cost and easy to install,and is suitable for partially occluded scenarios of key points and therefore has high robustness.It is verified that this method has high accuracy and can provide data support for extended research based on driver pose.The keypoint motion trajectory prediction method proposed in this paper provides the basis for the subsequent keypoint based driver pose prediction.