| With the continuous development of society, an aging population is already an inevitable trend. The elderly is the vulnerable group in our society, so they are very easily to meet various dangers in daily life. Moreover,many of them are the only child families,and their child often can’t accompany them because of the pace of life and work pressure,etc. so many of older people are in a unattended state. When the elderly are in dangerous because of falls or other reasons, they are often unable to obtain timely relief and caused major injury. Thus, it’s very necessary for the elderly to create a smart home monitoring system, which can monitor and analyze their real-time behavior and automatic detect the falls and other abnormal actions, so that the elderly can get a timely assistance when they are needed.Currently, Although there have some methods for detecting the elderly’s daily behavior activities, they are many insufficient to meet the practical application. The most popular method is the behavior detection based on wearable devices, the elderly need to carry a variety of sensing devices in their body, this problem limit its application greatly. The another is the behavior detection based on daily activity information, this method get the elderly’s daily activity information by installing many sensors on the home, then create the person’s activity model,and detect the abnormal behavior by this.it depends on many kinds of special sensors heavily. Due to the continuous development of smart surveillance technology and its great usability, the behavior detection based on computer vision get the more and more attention, In this paper, we carry out the research about the abnormal behavior recognition,which is based on computer vision algorithms, the main research work was done:(1) A shadow elimination algorithm based on HSV color space and texture feature is put forward. According to the HSV color space, detecting the shadow region and removing it,and getting the body’s candidate region, then abstracting the gradient spatial information to convey the image’s texture as a supplement. Finally, removing the image noise by median filtering and morphological processing, and making it as the next input image for feature extraction.(2) Adaptive from Spatial moment and projection histogram feature to abstract the human body’s shape features and statistical features. The human body’s motion feature will be expressed by computing the Euclidean distance about the centroid of adjacent two images. Finally,we combine the three characteristics as the most effective description of the behavior,which contains spatial information and time information.(3) For the sake of validating the overall performance of the system,we establish the behavior video database.Videos consist of multiple types of human actions are collected by a Canon IXUS115HS camera in the indoor environment. The activity database set called OwnDataBase,which is constructed by the collected videos. The OwnDataBase contains140video clips, involving six daily behavior, which are walking, jogging, sitting down, squatting down, bending and falling down (forward and backward). The experimental environment match the real scene to some extent.(4) Through the study of the existing classification algorithms, we propose a classification framework for the abnormal detection, which make support vector machine as the core classification algorithm. According to the abstracted behavior features,we can obtain the unknown behavior class label,and distinguish from the abnormal behavior. Experimental results show that the proposed algorithms can detect and identify the fall or other abnormal behavior from multiple daily activities,and the normal behavior can be also identified with good recognition performance. |
PDF Full Text Request