| Sitting posture recognition is a key subject in the analysis of human behavior,and it has wide application prospect in military,education,monitoring,medical and other fields.Compared with the traditional sensor-based sitting recognition method,the visual sitting posture recognition method is based on the sitting image analysis,and has the advantages of simple implementation,rich information and more intelligent deployment.Particularly,the advance of depth camera greatly simplifies the threedimensional modeling of the human body sitting posture,which is conducive in improving the visual sitting posture recognition reliability and accuracy.Based on this,this paper mainly studies the sitting posture recognition based on depth image.This paper aims at accurate sitting posture identification and the application of sitting posture recognition,it mainly focuses on the acquisition and processing of the sitting posture depth image,the feature extraction of the image,the accurate identification of the sitting posture and the design of the sitting detection system.The main work and innovation of the paper are as follows:(1)A fast and effective image segmentation algorithm is designed for sitting posture.In this paper,based on the threshold segmentation method,the actual depth distance is used as the segmentation value,and a threshold updating method is designed according to the characteristics of the sitting posture,so as to realize the effective segmentation of the background image.In addition,the external box update method is designed to realize the removal of external interference,taking into account the interference situation encountered in the sitting situation.Experimental results show that the segmentation algorithm can effectively segment the human body in the depth image and has good real-time performance.(2)A sitting posture recognition algorithm based on key point is proposed for sitting posture in simple scene.Firstly,the posture key model is designed by combining the "Vitruvius".Then,the head of the body is located based on 2D contour mean square template matching algorithm,and the shoulder and body center are positioned by using the characteristics of human body,contour feature and depth.Finally,the characteristics of the angle,the depth of the key points and the contours of the images are extracted to obtain the sitting judgment criterion.The sitting posture in a simple scene is identified through this criterion,the average recognition rate can reach 95.92%,and the average time of recognizing an image is 50 ms,which has good accuracy and real-time performance.(3)A sitting posture recognition algorithm based on depth image projection is proposed for the sitting posture in complex scene.Firstly,the projection depth image is projected to obtain the projection image,and the projection image is blankly removed,bi-cubic interpolation scaling,normalization and PCA dimension reduction to obtain the projection feature.Then,the foreground view is extracted from the pyramid HOG feature and the HOG feature of the depth map is obtained.Finally,the sitting posture feature vectors are obtained by combining them,and the random forest classification algorithm is used to train and classify.The final test results show that the algorithm can achieve 99.27% of the average sitting posture recognition rate in complex scenes.(4)A system for sitting posture detection based on the Android platform is designed and implemented.The environment of the sitting detection system is established,and the sitting posture recognition algorithm is transplanted and realized on the Android platform.Then,based on the posture recognition algorithm,a number of practical functions such as sitting detection,image display,sitting reminder,sitting statistics,state analysis,database storage and image playback are designed for the sitting detection system.Finally,the effectiveness and reliability of the system are verified by the actual application test. |
