A Simple Fall Detection Scheme for Early Detection of Falls in Elderly People
Srijita Ghatak, Rishita Mitra, Washef Ahmed, Kunal Chanda · 2023
Human fall is considered to be a prominent risk especially among the elderly population and for disabled people. Fall detection systems often rely on video surveillance, capturing frames and utilizing various algorithms to identify instances of fall. The continued development of such systems is one of the goals of AI technology, which aims to improve quality of life and reduce fall risks. The field of computer vision still faces significant difficulties with the estimate of human postures. Recognizing the poses of multiple individuals in real-world scenarios presents greater challenges compared to recognizing the pose of a single person in an image. An automated camera-based system that can recognize abnormal human postures is presented in this work. To identify people in frames and extract their body key points, a pre-trained pose estimation algorithms like PoseNet and OpenPose is used. The human body geometry that was recorded at various frames of the video series is used in this paper's proposal of a fall detection method. To generate the active energy image, the average of the difference frames obtained from the silhouette sequence is calculated. The resulting image is then divided into multiple segments. For each section, affine moment invariants as gait features are computed. This paper will provide a detailed description of the fall detection technique. Results show an accuracy of more than 90%. On the test set, this method had an F1 score of 92.5%, accuracy of 92.5%, precision of 91.2%, recall of 93.8%.