Comparative Analysis Of Deep Learning Models For Human Fall Detection Using Skeleton Features From Video Surveillance System

Detecting human falls is a critical task within video surveillance systems, as it can generate an alert to caregivers during the event of a fall, leading to major bone injuries and might cause death. Vision-based method is an approach that can be used for detecting human falls from video surveillance system. In vision-based method, the use of skeleton features, which are the positions of the joints in human body can serve as primary attribute for fall detection. The proposed method involves multiple steps. First, datasets for fall detection, such as UP and UR, are prepared to facilitate the training of deep learning models. In order to extract skeleton features from input data, Mediapipe framework was utilised. The skeleton features are then used to train deep learning models which then classifies the movement as fall or non fall. In order to assess the effectiveness of the models, the UMR fall detection dataset was employed for performance evaluation. In the proposed method, various deep learning models namely LSTM, CNN-LSTM, TCN, CNN as well as machine learning model like Random Forest are trained on extracted skeleton features. According to the results, the LSTM model performed well on detecting Fall and Non-fall activities with test accuracy of 80% compared to other trained models. However, one drawback of LSTM model was its inability to distinguish between fall and sleeping activities. The findings demonstrated that Random Forest was capable of distinguishing fall and sleeping activities. Overall, the proposed method presents a comparative analysis of various deep learning and machine learning models to detect human fall from video surveillance system, with skeleton features extracted using Mediapipe framework.