With the advent of population ageing and the problems of the empty-nest elderly, more and more attention were paid to the elderly. It has been discovered that fall is a fatal reason of injury related death, besides, the occurrence of injury is relatively frequent and the damage is relatively serious. In order to improve the quality of life of the elderly, as well as, to reduce the injury of their body and property damage, in this paper we study the fall behavior in detail.In general, the research methods of fall can be divided into three categories, that is, wearable device based method, ambience sensor based method, camera (vision) based method. The recognition rate of the first two methods are relatively low, and they depend, to a great extent, on environment. However, the vision based method have a relatively high recognition rate and it can detect multiple events at the same time, more important, it can record the video for postmortem analysis. So in this paper we chose the vision based method to study the behavior of fall.Firstly, in this paper the moving object had been extracted based on background difference method. In general, the initial background frame was established simply using the first frame of the video, it takes more time to approach the real background. We improve the traditional methods by establishing the initial background using the median method. Again, the background had been established and updated by Gaussian Mixture Model (GMM) and the foreground object was distinguished from the background by matching algorithm.Secondly, using the improved HSV color space detection operator to eliminate the shadow solving the problem of traditional method, which is hard to select parameters. Furthermore, the gradient feature were extracted to make up the shortfall of HSV color space based method. Because the object boundaries extracted by the traditional gradient operator is not complete, in this paper, four directions,0° ,45° ,90° ,135 °, were chose to detect the edge of the object. Combining the above two method to get a more complete object.Then, the shape feature, which has the invariance on translation, rotation and scaling, Hu invariant moments were extracted. The first four Hu invariant moments, which have small amount of calculation were chose. Combined with aspect ratio, attitude rate and velocity together to make up 7-dimensional vectors as the feature vector to describe the current frame. We extracted 15 frames by interval sampling as a motion cycle, that is,105 dimensions feature data together to describe a behavior.Finally, the six behaviors which is walking, jogging, sitting down, squat, bend, falling, were classified by Support Vector Machine (SVM). The six behaviors were firstly divided into two categories:upright state and non-upright state. For the two behaviors which belongs to the upright state, only one SVM is needed. For the other four behaviors which belongs to the non-upright state, six SVMs is needed. After such a grouping, the number of SVMs was reduced, as well as, the classification rate was improved.In this paper the index of sensitivity and specificity were used to evaluate the classification results. The higher the sensitivity, the lower the missing rate. The higher the specificity, the lower the false positive rate. It had been verified that the proposed method can effectively identify the six kinds of behavior, with a relatively high sensitivity and specificity. According to the statistics, the average correct identification rate of the sample is 92%. |