Pose Estimation based Fall Detection system using MediaPipe

Vision based Fall detection systems are much better in terms of convenience and accuracy over other Fall detection methodologies which rely on additional sensory data. The computational cost associated with large number of parameters within non vision based Fall Detection systems make them inefficient performance wise in real time. Feature extraction from video sequences is a very extensive task when done through conventional methodologies and proves to be highly computationally expensive. Due to the rising number of falls in the world and its consequences, a need to develop a fall detection system has emerged, which is not computationally expensive and provides a real time detection with least false positives. A pose estimation based approach which utilises the skeleton key points of human body with help of MediaPipe is explored. Decision conditions are used to classify whether a fall has occurred or not, that include - the velocity of descent over frames, decrease in height trend of the bounding box and the bounding box width-height ratio change over frames. The accuracy obtained is 95.84%.