The Study On Machine Vision-based Indoor Fall Detection

Accidental fall is one of the major health threats to the elderly population which can cause serious physical and psychological harm or even be fatal. With the serious aging problems, it is significant to detect falls for the elderly and it has become a hot topic. Although different methods have been proposed in this field, there are still some problems to be solved in distinguishing falls from daily activities. The fall detection methods based on machine vision provide an effective way to monitor whether the elderly fall with the help of the gradually popularization of the monitoring facility.The aspect ratio of bounding box and centroid of human body silhouette(HBS) are two typical features in this field, however, the former is easily influenced by human limbs, and the latter is difficult to distinguish falls and fall-like activities because of the changes resulting from centroid. In addition, both of them have limited ability to detect the falls parallel to optical axis of camera. To address above problems, this paper proposed a new multi-centroid motion vector(MCMV) feature as well as the extraction method. Meanwhile, the proposed feature is applied to two novel fall indoor detection models based on machine vision. The main contributions of this paper are as follows:This paper employs Mixture Gaussian Background Model to extract HBS, which is fitted by an approximate ellipse with removing the region outside the ellipse to locate the main silhouette of human. The above main silhouette is divided into three regions according to the aspect ratio of HBS bounding box followed by tracking the centroid of each region and extracting their changes as MCMV. K-means clustering algorithm is applied to analyze MCMV to produce vision words for founding Bag-of-Visterms(BOV), the human motion is transformed into vision statements, as well as the complex activities problems are decomposed into relatively simple text analysis problems.According to the characteristics of vision statements, this paper proposed two indoor fall detection models based on machine vision. The vision statements with different length are transformed into the feature vector with the same dimension by term frequency. The proposed directed acyclic graph support vector machine(DAG-SVM) serves as the fall detection model with above feature vector as the input. In order to make use of the continuity of human activity, this paper proposed a fall detection model based on Hidden Markov Model(HMM), in which vision words and vision statements are regard as the observation state and observation sequence, respectively. This model analyzes the spatial-temporal process of human motion to determine whether the body falls. Simulated experiments showed that the two proposed methods have achieved a high accuracy in fall detection.