An optimized edge AI framework for real time fall detectionwith elderly care using quantized yolov8-pose

Abstract

In particular, the timely intervention of medical personnel when there are falls of elderly patients plays an important role in lowering the rates of morbidity and mortality. Although visual-based deep neural networks are capable of delivering precise classification results regarding human action recognition, the use of these algorithms in low-power edge devices is not possible due to issues like latency and high thermal signatures. We propose a privacy-preserving Edge-AI approach for fall detection of humans in elderly care applications. The proposed Edge-AI framework utilizes a one-stage YOLOv8-Pose algorithm for extracting anatomical joint structures. In order to aid execution on constrained resources of an embedded microprocessor, post-training asymmetric quantization compresses neural network parameters from 32-bit floating point (FP32) format to energy-efficient 8-bit integer (INT8) format. The kinematics classifier downstream detects two different spatial and temporal features: the structural displacement ratio of the bounding box of the human subject and the instantaneous vertical velocity of the pelvic center of mass. With the help of local inference, there is no dependency on the bandwidth of cloud services and complete user privacy. The model being developed strives for reducing physical storage sizes by 75% along with a frame rate of 30 fps on commodity hardware.