Estimation of Human Fall Detectionusing Body Pose Rule based Algorithm for Video Surveillance

Human fall detection is hot topic in the field of surveillance and health. The proposed approach consists of two parts: object detection with tracking and fall detection confirmation. The novelty of our work is that we have implemented algorithm which is composed of two rules for human fall detection. The outcomes of the algorithm demonstrate how correctly our model can identify the majority of single or multiple human falls quite accurately. The algorithm is performed on the publicly available benchmark UR Fall Detection datasetand analyzed using evaluation performance metrics. This paper aims to detect falls in a video stream using a combination of YOLOv3MobileNet model for human detection and AlphaPoseResNet model for detecting human poses and their keypoints.YOLOv3MobileNet pretrained model reduces the computational complexity and beneficial for real time fall detection. AlphaPoseResNet pretrained model having deeper backbone network offering high accuracy in pose estimation. One of the main challenges in fall detection systems is minimizing false positives while maintaining accurate fall detection. The proposed method shows that false alarms can be minimized while enhancing detection accuracy.