Measurement And Analysis Of Human Body Posture On Car Seats

At present,the measurement of 3-D human body posture on a car has a crucial effect on the comfort design of a car seat.Traditional manual measurement not only has a low degree of automation,a large investment of manpower and material resources,but also has poor accuracy of measured data.Based on deep learning technology,this dissertation studies the method of combining binocular vision and deep neural network,which can improve the speed of 3-D pose measurement and the accuracy of measurement data.Firstly,through the investigation of the traditional human body attitude measurement methods and the existing 3-D human body attitude measurement methods,considering the complexity of the measurement scene,equipment manufacturing cost and measurement accuracy,the binocular structured light measurement system is selected as the 3-D human body attitude data acquisition equipment in this dissertation.According to the requirements of the light of the scene to be measured,the size of the object to be measured and the measurement accuracy,the configuration of the camera,camera lens,structured light fringe projection device and data processing device required by the system is selected,and the construction of the binocular structured light measurement system is completed.Secondly,after investigating the existing 3-D human posture measurement technology at home and abroad,this dissertation proposes a 3-D human posture measurement method based on depth neural network.This method combines the extraction method of 2-D human joint depth network with the binocular measurement system,takes the output of the first 10 layers of the improved VGGNet-19 network as the input of the depth learning network,and uses the two-channel multi-stage iterative network to extract the 2-D human j oint and limb positions respectively.Combined with the brief feature of joint position and the epipolar constraint of binocular camera,the matching information of 2-D joint points collected by left and right cameras is obtained.The matching 2-D joint point information is transformed into 3-D space by binocular camera calibration results,and finally the 3-D human posture information is obtained.In addition,the three wavelength structured light algorithm of λ1’=λ123=1008,λ2’=λ23=144,λ3’=λ3=16,is selected to obtain the complete 3-D data of human body.Compared with the traditional multi frequency heterodyne 3-D reconstruction algorithm,it can effectively avoid the point cloud loss caused by mismatching.Provide more detailed parameters for car seat design,Finally,in order to verify the effectiveness of the algorithm proposed in this dissertation,the system is calibrated and used to extract the 3-D joints of the human body on the vehicle.The experimental results show that the calibration error of the binocular system is 0.33mm,while the joint detection rate of the deep neural network for the complex scene in the COCO2017 dataset is 75.3%,and the detection rate in the self-collected data set is 92%.By using epipolar constraint and brief feature to complete the missing joint,the detection rate of the system can be improved to 98%,and the stability of the key pose angle calculated by obtaining the three-dimensional joint point is less than 10°.The experimental results show that the proposed method can meet the data acquisition requirements of the actual car seat design,and can also be extended to other 3-D attitude measurement requirements.