Fall Detection Based On Temporal And Spatial Variation Of Human Body Posture

With the development of intelligent video surveillance technology, the detections and analyses of human behavior have been promoted. Home security is an important part of these studies. The use of video analysis on detection of falls and other accidents for the old people in the home is undoubtedly of great significance.In this paper, we firstly analyze the human detection methods in video surveillance. Considering that the HOG feature has large calculation and poor real-time results, we select the CENTRIST algorithm to detect pedestrians in video. We also transplant the CENTRIST algorithm on Hisilicon Hi3516 platform.There are non-erect human body postures in fall events. The variability of human gesture will lead to the increase of mistake rate and miss rate. This paper study on the body posture Poselet detection method. Poselet is a particular posture of human body. We get training examples of Poselet by key point annotation and body part similarity measure, then we extract the HOG features of the training examples and train them by linear SVM classifiers. We use these classifiers to detect the specific posture. We expand the Poselet samples for fall detection to improve the variability of the original Poselet example dataset and make it more suitable for fall detection. We train Poselet classifier under different angles. The experimental results show that our Poselet classifiers can detect the human body under special postures, such as bending, lying, which is the basis of the fall event detection.Considering that the body posture will change in both time domain and space domain, such as the relative position changing of head and torso. This paper study on the spatial positions and key part angles of human body at different poses and construct a state machine model to represent the conversion between different postures. We propose a posture and fall event detection method based on temporal and spatial changes of body parts. The experimental results show that our method has good effect to resolve different posture and detect the changes of human body posture, thereby detecting the occurrence of fall events.