Research And Implementation Of Human Fall Monitoring System Based On Raspberry Pie

Nowadays,the effective life span of human beings is increasing,and more and more elders live alone.In these contexts,falls are slowly becoming the number one cause of accidental death among old people who live alone.How to solve the problem of monitoring fall for elders and automatic help-seeking at low cost has very high practical significance.There are three main research methods to detect fall in China and abroad,that is,werable device based method,ambience sensor based method,computer vision based method.Because of the convenience and the none of invasive,computer vision based method of fall detection has a great future.So in this paper we chose the computer vision based method to detect fall.The processing of the fall detection system mainly included three steps: moving target extraction,feature calculation,gesture and motion recognition.In the step of moving target extraction,this paper mainly used Gaussian Mixture Background Modeling.In order to solve the problem that the foreground object disappears when it remains stationary for a long time,this paper proposed a method of fusion edge extraction and background difference.It can improve the extraction effect.For edge extraction algorithm,this paper studied Sobel algorithm and introduced the 45 ° and 135 ° filtering template that improve the extraction behavior.In addition,we used HSV color space to eliminate the shadow.Finally,we used morphology process,gaussian filtering,median filtering and other operations to extract the foreground target with clear edge,substantial interior and high usability.In the step of feature extraction,this paper extracted various static and dynamic features.Before feature extraction,the region of interest needed to be marked.On the basis of the common method of extracting the minimum enclosing rectangle,we gave example of deficiencies and failures.And we introduced a new method to extract the minimum enclosing rectangle that contrast between the distance from the center point and outline points.It can resist the background noise and ensure the quality of foreground extraction.And then we designed the method to extract other features such as aspect ratio,eccentricity,centroid coordinates,Hu moment,fourier descriptors and so on.By combining features into vectors,we completed the transformation from image to numerical space.Finally,we obtained the optimal feature combination through feedback experiment.In the step of gesture recognition,in order to detect the indoor falling behavior better,this paper collected several common indoor behaviors as experimental data sets,including walking,sitting,squatting,bending,and falling.We introduced support vector machine to identify action on single-frame image and designed its experiment.we devided the video into frames and divided them into training set and testing set.After determining the kernel function and completing parameter search,the feature vectors of moving targets in each frame were trained and classified,and the final average recognition accuracy is 98.72%.On the other hand,we introduced hidden markov model to identify action on multi-frame images and designed its experiment.After obtaining the feature vector of each frame,we clusted them so that the vector combination of a complete action sequence can be reduced to a label sequence of finite observation states,which can be used to model by hmm.The final average recognition accuracy is 87.01%.Raspberry pie is an ideal design platform with low cost,low operation cost and low development cost.Based on the above experimental results,we designed a fall monitoring system based on raspberry pie.The system is divided into two parts: server and client.The server is based on Raspberry Pie and CSI camera,and has multi-threaded service capability.The client can run on general-purpose computer of Linux,Windows and Mac system,and it provide GUI operation interface.