Research On Unmarked Cervical Motion Detection Method

As society advances,people’s quality of life is improving,people are living and learning in different ways than ever before.College students,programmers,office workers and many other people keep bad neck posture for a long time when working with electronic products.If the bad posture exceeds a certain range of tolerance,then cervical spondylosis will become "younger".In recent years,the neck posture has attracted more and more attention.Based on the background of human body posture detection,this paper carries out research on neck posture detection and recognition,which has certain reference significance for the further study of biology and pathology.First of all,there are few studies on neck posture by referring to domestic and foreign literature.Common neck posture detection is to use acceleration sensor attached to the neck for marking to detect the angular velocity,linear velocity and tilt angle equivalents of neck movement,but the sensor attached to the human neck will cause discomfort.In order to avoid the above situation,this paper uses the methods of image processing and machine learning to detect and recognize neck movement so as to achieve the purpose of no mark.The experiment includes gray-scale,gaussian filtering and image segmentation for different neck motion images.Image segmentation adopts the least square fitting Otsu algorithm,which has the advantage of faster real-time processing speed compared with the traditional segmentation algorithm.Secondly,after the neck image segmentation is completed,feature pixels are extracted from different neck motion images through star model,Hu moment invariant,Zernike moment invariant and Wavelet moment invariant.The extracted feature pixel values are fused again by the differential evolution Gray Wolf optimization algorithm（GWODE）.Compared with other fusion methods,this algorithm not only reduces the dimension of neck image,but also improves the recognition rate of neck pose.Finally,after the fusion of feature pixel values of different algorithms is completed,each pose will have its corresponding feature vector.The linear classifier based on discriminant function method,support vector machine and cascade forward neural network are used to classify and recognize the feature vectors fused by different algorithms,and the accuracy of classification recognition is compared.Based on the classification accuracy and classification efficiency,the experimental result expresses the cascaded forward neural network is better for each feature fusion algorithm.