Indoor Human Fall Detection Based On Color-depth Images

The elderly fall has high occurrence and risk.Fall not only affects the health and safety of the elderly,and reduces their qualities of life,but also brings heavy medical burden to their families and the society.At present,China is becoming an aging society.The elderly population is growing rapidly,and the number of empty nest elderly is rising as well.The occurrence of fall in the elderly is increasing.With the development of visual sensors and computer vision,the human body fall detection study based on image-based method has obtained much concern,and it is significant and valuable to use image-based equipments for human fall detection.The fall detection studies used to utilize the color images or depth images.However,there are some problems with these methods,such as insufficient information used,or the poor performance in various illuminating environments.In addition,former studies focused on the detection of moving objects and thus extracted only the foreground features.These foreground features are relatively sipmle,and are not representative enough.In this study,based on the Kinect color images and depth images,we firstly proposed a machine learning method by extracting features from color-depth images to detect fall.Then,we combined our method with moving objects detection methods to promote the performance of the fall detection.This studies mainly include the following four parts:Firstly,we proposed to use the regional image features in for fall detection.We analyzed the characteristics of non-fall and fall images,and found that in the fall images the features,such as color,texture,shape or spatial features,are not stable;however,in the non-fall images,the head-shoulders have the same shape and similar spatial relationships.Therefore,we used the head-shoulder features for fall detection.Secondly,we proposed to combine the color images with depth images for fall detection.Regarding the regional image feature extraction,we calculated the histogram of oriented gradient(HOG)to extract features from color images,and two depth descriptors—sorted depth self-similarity(SDSS)and PCA-DSS in the depth images.Finally,we combined these features to detect falls by using support vector machine(SVM).The depth image features were consistent in various illuminating environments,thus our method not only improved the detection accuracy,but also extended the application fall detection in different environments.Thirdly,based on the above methods,we combined with the moving objects detection for fall detection.We extracted the height(H),width(H)and H/W features from the human foreground image,and applied the contour of human body ellipse to extract the tilt angle as well.Then,we combined all these features with regional image feature detection to detect fall by setting thresholds for improving the performance in fall detection.Fourthly,we tested and verified the performance the above proposed methods in various illuminating environments.We conducted a series of experiments to validate the fall detection algorithm based on the color-depth images.Eventually,it was found that our fall detection system achieved a high accuracy of 97.06% under normal illumination and 96.94% in bad illuminating environment.In summary,in this study,we proposed the fall detection methods by using the regional features in color-depth images for fall detection,and introduced two new feature descriptors in depth images.By combining the moving objectives detection method,we achieved a relatively high accuracy in the human fall detection both in the normal and bad illuminating environment.Our proposed methods show substantial innovativeness in human fall detection,and are supposed to provide solutions for the elderly fall which is a big problem for the families and the society with high occurrence and high risk.