Computer vision based human fall detection and classification for real-time videos

Abstract

Most recently, Human fall detection systems using deep learning models find major applications in all fields, especially in the held of healthcare. Even without doctor analysis, most Neurological and musculoskeletal diseases such as oncoming strokes and gait problems can be identified using these models and computer vision. In this article, automatic human fall detection is proposed using a convolutional neural network by applying real-time videos. In general, most of the research has been carried out using standard videos which will not apply to real-time applications. Hence this work concentrates about using convolutional neural networks as a system has real-time videos for the Human Fall Detection and monitoring system using three pre-trained models: (i) TinyYOLOv3-ones, (ii) AlphaPose and (iii) ST-GCN. The proposed Spatial temporal graph convolutional networks produce better accuracy with captured real-time video for human fall detection. The same method was also utilized for classification with different epochs. The results were compared and maximum accuracy of 100% is obtained for 500 epochs. Hence it is proved that the existing method can be utilized for human fall detection with greater accuracy.