Research Of Fall Detection Algorithm Based On Multi-feature Fusion

Nowadays, aging population has become a trend of globalization. In order to deal with the demographic change, countries all over the world have put great enthusiasm into elderly cared services and medical ward. Falling has been proven to be the second-most dangerous factor in accidents and unintentional injuries especially among the old people. Studies have shown that a large part of tragedies caused by falling accidents could be prevented or eased if the fall was found timely.In recent years, behavior identification technology obtained fast development. It is practically important and valuable to apply the behavior identification technology to the research of fall detection algorithm. The purpose of this paper is to study a fall detection algorithm which has stable operation and efficient performance. To overcome the poor description force by single feature, algorithm in this paper used the strategy of multi-feature fusion. This paper also proposed the modified algorithm to improve the effect of front-view target segmentation. The main work in this paper includes:At first, convert the video into an appropriate format and store them in grayscale image sequences. The images should have noise reduction.Secondly, study the classical algorithms of motion detection and choose the median method as the motion detection algorithm in this paper. In view of the front-view motion detection whose trajectory was covered, a modified algorithm is proposed. Using the fuzzy classifier to distinguish the sequence’s front and rear, deduct the background respectively. Perform statistical analysis of pixels without changing to complete target detection.Then, according to the main features of fall, multi-features are used to describe the behaviors. Motion target’s centroid changing rate is used to represent the motion features. When the frame’s centroid changing rate meet the threshold value, using the frames in the neighboring region together form VEI as the posture feature to give the quadratic discrimination. The improved SIFT algorithm is also be introduced as a local descriptor to give more details for occlusion. The improved SIFT introduce Harris corner detection in scale space to instead of original extrema detection. Before corner detection, count the similar pixels in neighboring region to screen the pixels that couldn’t be corners. The improved SIFT improves the operation speed and can adjust the number of corners by threshold.At last, this paper implements falling behavior classification by bi-directional feature matching on SIFT. Graphical user interface is designed by MATLAB to realize falling detection. The algorithm in this paper is tested on CASIA database and self-built database, their recognition rates reach 90% and 94% respectively.The detection algorithm in this paper turns to be stable and reliable. It can be applied for the detection of behaviors with partially occlusion. The algorithm’s recognition rates also get a promotion compared with algorithms in other papers. All the improvements mentioned above make efforts to the fall detection’s practical application.