Patient Monitoring System
Shubham Deshmukh, Favin Fernandes, Sakshi Kulkarni, Aniket Patil, Vaishali Jabade, Jyoti Madake · 2022 IEEE 2nd Mysore Sub Section International Conference (MysuruCon) · 2022
With the need to monitor the status of individuals, many researchers and scientists have used pose estimation techniques to record and analyze human poses. The research proposes a method to monitor patients based on human detection and human pose estimation and has proposed a methodology for it. The main purpose of this paper is to detect if a person in a given frame is a patient or visitor and then detect various poses of the patient only. For human detection, YOLOv4-tiny is trained on 2 classes: Patients and Visitors using the Google Colab platform, and the dataset was collected from open-source resources like Google images. Pose is estimated only for patients after training using the Mediapipe framework, which detects 33 keypoints. The model will classify a pose into 4 different classes: Sleeping, Sitting, Walking, and Standing. The keypoints from Mediapipe are then used to classify the pose of the patient, using PCA and Xgboost to monitor and export the data in a CSV file. After the training, the pretrained model is used to apply it in a real-time video application developed using the Tkinter GUI library in Python. The accuracy for the YOLOv4-tiny model is 91.09% and for the Xgboost model it is 98.82%.