Innovative System for Human Activity Recognition and Pose Estimation Using CSI WiFi and the YOLOv8 and Mediapipe Frameworks

Abstract

The present invention concerns an innovative system for human activity recognition and pose estimation using WiFi channel state information (CSI). This system enables accurate, real-time analysis of human activities and poses without the need for portable devices or intrusive cameras. Technical Description: The system consists of four main modules: 1. Data Collection: This module uses commercial WiFi devices to collect raw CSI data. Fluctuations in WiFi signals, caused by objects and movements in the environment, are captured using a WiFi router and a desktop computer. The data collected enables motion patterns to be detected. 2. Data pre-processing : The collected CSI data are pre-processed to improve their quality. This process includes antenna selection to optimize the signal, application of the Hampel filter to eliminate anomalies, Discrete Wavelet Transform (DWT) to reduce high-frequency noise, and Principal Component Analysis (PCA) to extract essential features and reduce the dimensionality of the data. 3. Human Activity Detection: The YOLOv8 model is used to analyze images associated with CSI packet identifiers. This model, trained to recognize activities such as sitting, standing, raising the hand, starting to sit and starting to stand, provides real-time detections with bounding boxes and labels for each recognized activity. The association of images with CSI package identifiers is carried out internally, guaranteeing data confidentiality. 4 Pose Estimation and Skeleton Extraction: The Mediapipe framework is used to estimate human pose and extract the skeleton from images associated with CSI packages. Mediapipe detects key points and builds a skeletal representation of the observed individual, providing detailed information on posture and movements. This approach improves the accuracy and robustness of human activity recognition.