Research On Fall Detection Based On Surveillance Video

With the development of society,the trend of aging is becoming more and more serious.When the elderly are alone,the fall cannot be discovered and rescued in time,which may cause serious injury or even death.At present,fall detection research has become one of the important research topics.However,due to the non-deterministic decline,illumination,occlusion,cost and other factors,it is of great practical significance to study and design a low-cost and high-accuracy fall detection system,this paper completed the research of the fall event detection based on surveillance video and the design of the prototype system for the detection of fall events using only one surveillance camera in each room.The quality of the fall detection is usually limited by the camera's perspective or the occlusion of objects in the scene.This paper proposes a method based on contour deformation,which matches the similar parts of the human contour between two consecutive images.Focusing only on the unobstructed part to quantify the deformation,thus enhancing the robustness of the system to occlusion.First,the contour of the body is obtained by foreground segmentation,and the edge points are extracted from the contour by the Canny edge detector,and the body deformation is analyzed by comparing two continuous contours.Secondly,the shape context matching is performed,the contour edge points are matched with the video sequence,shape analysis is performed by Procrustes distance to obtain fall detection criterion.Finally,the Gaussian mixture model(GMM)is used for the fall event detection,and relevant experimental analysis was carried out,and compared with the traditional method.Because there is an abnormal event in the fall detection system based on 2D surveillance video,in order to improve the detection accuracy of system falls,a method of recovering 3D head information based on 2D image and tracking it to complete the detection of falls is proposed.In the system,only one calibration camera is used.The 3D head model is projected into the image plane through the camera's internal and external parameters,and the ellipsoid information of the head is obtained,The head is tracked by four-layer particle filter.The head vertical velocity Vv and the height of the ellipsoid from the ground are used as features to detect the fall events.The detection threshold of the vertical velocity Vv of the head is set to-1m/s,and the fall occurs when the vertical velocity Vv of the head is less than-1m/s;head height threshold is set to 50cm,and falls are considered when head height is less than 50cm.Through experimental analysis,the fall accuracy rate was 93.67%,the specificity was 93.3%,and the sensitivity was 94.34%.Improved the accuracy of fall recognition compared to 2D systems.At the end of the paper,in the case of using only one surveillance camera in each room,the prototype system based on monitoring video fall detection is designed to complete the detection of fall events,reducing the cost of fall detection and facilitating the application of fall detection technology.