Research And Realization Of Video-based Automatic Fall Detection

Falls in elderly are gradually becoming an important public health problem. Falls may cause unconsciousness, disability or even death among old people. So, it is necessary to track and analyze the actions of the aged, and to get their fall actions correctly detected to give them timely aid.There are still some deficiencies in the current fall detection methods. The wearable fall detection method, for instance, is greatly restricted in its actual application for the inconvenience of carrying sensor devices. The ambience device method, which uses pressure sensors to detect the emergence of the human body and get its position, can’t recognize whether the pressure really comes from the user’s weight or not. While the video-based fall detection method has got more and more attentions due to the ubiquity and the practicability of video. This thesis also makes investigations into the video-based automatic fall detection algorithm, and its main investigations are as follows:1. Thorough investigations and researches are made into the moving target relevant to the video-based automatic fall detection algorithm. The moving target will be detected, tracked, segmented and image preprocessed in the thesis. This thesis, combining the threshold method and the Gaussian Mixture Model method, makes moving targets segmentation, removes the shadow on the color information based shadow suppression method, removes the noise using inflation and corrosion operators, and finally get a more complete segmentation image of the target. Experimental results show that the algorithm proposed in the thesis can make moving targets segmentation better.2. Through the analysis of the current fall detection algorithms, this thesis proposes an improved video-based automatic fall detection algorithm, which uses characteristics of human aspect ratio to carry out the fall detection. The advantage of the improved algorithm is that characteristics of human aspect ratio are easy to analyze and therefore the complexity of the algorithm will be greatly decreased and the calculation work greatly reduced. But in some special cases like sports actions, some sudden squat actions or getting up action after the fall, misjudgments will appear if only by using aspect ratio. Two additional judgment conditions--effective area ratio and center variation rate--are thus employed into the algorithm, which successfully modifies the above misjudgments. Experimental results show that the improved algorithm has its feasibility.3. A video-based automatic fall detection system is developed, which has integrated the improved algorithm. Experiments of this thesis indicate that such a system can detect fall actions and have practicability.