A Vision-Based Abnormal Behavior Detection System For Elderly

Aiming at the problem of old age,the intelligent video surveillance system realizes the daily monitoring of the elderly living alone indoors.However,most of the current video surveillance systems only detect falls for the elderly.This thesis designs a detection system containing multiple abnormal behaviors of falls,abnormal neck state,and long-time static.The system has a complete structure and further improves the abnormal behavior detection system.This system consists of data acquisition,behavior analysis,and remote alarm.The data acquisition part consists of embedded devices that transmit the images collected by multiple cameras to the behavior analysis part.The behavior analysis consists part of local experimental devices that undertake the main algorithms to detect various abnormal behaviors.The remote alarm part transmits the occurring abnormal behaviors to the cloud platform through NB-Io T(Narrow Band Internet of Things)technology.The system includes four modes: single-room single-camera,single-room multicamera,multi-room single-camera,and multi-room multi-camera to meet different needs.The fall detection detects the human body bounding box by the YOLOv5 s target detection model.A new target tracker based on the extended Kalman filter is designed to solve the problem that YOLOv5 s could not detect the human body with a special shape.Then extracts the features according to the detected human body bounding box.Putting the values of the features into the designed fuzzy logic system.The curve of the output value of the fuzzy logic system under normal behavior was analyzed.The falling threshold was set to 0.67.The abnormal neck state detection includes low-head state and neck tilt state detection.The Lightweight Open Pose human posture estimation model extracts the skeleton information of the human body.The low-head state index is represented by the ratio relationship between vectors.To reduce the error caused by direct detection,the neck tilt state is determined by camera calibration and calculation of the Euler angle to determine the human body orientation first,and then calculate the angle between the neck and shoulder after an affine transformation.The long-time static detection proposed the HOG(Histogram of Orientation Gradient)static detection and the frame-difference method static detection.The HOG static detection process includes pre-processing,HOG value calculation,and stationary index calculation steps.The frame-difference method static detection includes obtaining a frame-difference map,morphological processing,and other steps.Both methods are feasible for long-time static detection.The results of testing each abnormal behavior of the system showed that the accuracy of fall detection was 94.53%.The accuracy of head-down detection was 100% when facing the camera.The larger the rotation angle facing the camera,the lower the accuracy.For neck tilt state detection,when the human body rotated 20° to the right,the average absolute value of the relative error of the processed data was reduced by 11.38%.The long-time static detection of the Temporal Difference static detection was better than HOG static detection in detection effect and real-time.Finally,each mode of the system was tested.The detection results under each mode and the display effect on the cloud platform were shown.