Unifying Human Pose Estimation in the Fall Detection Problem

Abstract The skeleton model is extensively used to represent the human body in action recognition systems. Currently, numerous software and hardware solutions are available for building various skeletal models. The most well-known include Microsoft Kinect, Intel RealSense with Cubemos SDK (3D skeletal model), AlphaPos, PoseNet, OpenPose, and YOLOv8 (2D skeletal model). Typically, human action recognition systems are designed for one of these models, leading to system dependency on specific hardware and software. This dependency complicates the training of neural networks across multiple databases simultaneously for human activity monitoring applications. To address these challenges, we propose a uniform skeletal model and methods for transformation of various skeletal model types into that model. This approach aims to mitigate compatibility and interoperability issues in heterogeneous computing environments. Since skeletal models can be either 2D or 3D, their integration into action recognition systems can be complex. To simplify this, we propose activity map tuning for 2D skeletal models. Experimental studies confirm the effectiveness of the proposed methods for transformation of skeletal models and tuning activity maps. The results show that the integrated model not only maintains accuracy but also facilitates the development of more flexible and efficient systems for human action monitoring, potentially leading to broader applications and improved performance in real-world scenarios.