Human Fall Prediction by Gabor Filtering and Posture-based Feature using MediaPipe

This paper introduces the Human Fall Prediction Model using Gabor filtering and Posture Feature (HFGP). The proposed HFGP composes of Gabor filter feature extraction, angle calculation from human skeleton extraction by Mediapipe and classification by GRU model. The angle is computed by three points from human skeleton extraction by Mediapipe. These three points consist of midpoint of shoulder between right-side and left-side shoulder, midpoint of hip between right-side and left-side hip, and ground point. For experiments, the URFall detection data-set is used. The proposed HFGP model with a GRU architecture achieved an accuracy of 0.9838, outperforming both the LSTM and RNN models.