Fall Detection Based On Multi-feature Fusion Of Human Head And Body And Deep Learning

With the aggravation of population aging,the health problems of the elderly have attracted more and more attention.Falls is a fatal danger for the elderly,especially for empty nesters.If the elderly fall and fail to rescue in time,it may cause life danger.Therefore,it is of great value to develop a fall detection system,which can accurately detect the falls of the elderly and issue alarms.In recent years,with the development of computer vision,the field of video-based behavior recognition technology has also been very rapid development,there are many advanced algorithms can accurately classify all kinds of behavior.As a kind of abnormal behavior,if these techniques are applied to the detection of falling behavior,they will have important scientific research value and practical significance.The purpose of this paper is to study some problems in traditional fall detection methods based on geometric features.First,the geometric features are unstable,and some similar activities can not be recognized.Most of these algorithms use only one global geometric shape to represent pedestrians,and it is difficult to extract representative features to classify some similar activities as falls.Second,in the fall classification method,the accuracy of using traditional data statistics and linear discriminant method is not high.At present,some methods using in-depth learning have high accuracy,but the model is complex and needs a lot of training time,so it is difficult to meet the real-time requirements.Therefore,this paper proposes a method of fall detection based on multi-feature fusion and deep learning of human head and torso.In view of the instability of geometric features,it is difficult to distinguish similar activities.In this paper,a method of extracting geometric features of head and trunk separately is proposed.Firstly,Gauss mixture model is used to detect the foreground,the head position is predefined by a certain proportion,then the head is tracked by mean shift method,and the head is segmented by modifying the area size and position of the head.Then,the head and torso are fitted by modifying thetraditional ellipse fitting method,and three ellipses are extracted from the head and trunk respectively.Circular feature: The ratio of the long axis to the short axis,the direction angle and the velocity of the center of mass of an ellipse are fused into a space-time feature to express motion.Aiming at the accuracy and real-time classification of falls.In this paper,a shallow convolution neural network model is introduced to classify spatiotemporal feature sequences.The model is simple and consists of single convolution layer,pooling layer and full connection layer.In this paper,the extracted spatio-temporal feature sequences are used as samples to train and classify the model.Compared with the traditional classification methods,this classification method uses deep learning,which can effectively improve the accuracy.Compared with the existing deep learning methods,it greatly reduces the training time and meets the real-time requirements.In this paper,the above methods are tested comprehensively.The proposed model is trained and tested by capturing indoor scene videos and collecting a large number of temporal and spatial features.Compared with many classical fall detection algorithms based on geometric features,it is concluded that the proposed algorithm can effectively overcome the instability of geometric features and accurately distinguish some similar activities.For indoor places such as families,hospitals and nursing homes,the algorithm can be used to effectively protect the health of the elderly,and has important social value.