''' This repository represents my undergraduate thesis. I have introduced a novel approach for Human Fall detection in videos using two stream three dimensional(3D) Convolutional Neural Network and the evaluation is performed on two of the publicly avialable Le2i and URFD Dataset. '''
Title: "A TWO STREAM FUSION NETWORK FOR HUMAN FALL DETECTION IN VIDEO SURVEILLANCE"
Abstract
Falls are a major concern for the elderly population and can lead to severe injuries or even death. Substantial funds are reserved for the treatment of post-fall injuries and emergency assistance. Fall risks and their consequences would be significantly minimized if a fall could be reliably predicted or detected on time and prevented by offering prompt assistance. Thus, there is a need for substantial study and development of fall detection methodologies. To provide a proper response, fall detection technologies should be dependable and efficient. Several approaches for preventing or predicting falls in the elderly have been proposed. In this study, a fall detection methodology is introduced for the elderly using a two-stream three dimensional (3D) CNN network, which integrates visual and temporal information from surveillance video. The methodology can be applied to address the challenges posed by real-world surveillance scenarios by using a two-stream fusion architecture that processes both spatial and temporal information. Experiments on two publicly available dataset of real-world surveillance videos show that the proposed method outperforms state-of-the-art fall detection methods in terms of accuracy and computational efficiency, demonstrating the effectiveness of fusing the two networks' outputs. The results show that the proposed approach achieves an accuracy of 93.87%, specificity of 87.5% and precision of 89.0%. The suggested method has the potential to aid in the timely delivery of medical treatment to the wounded, and is therefore an important resource for public health.
Proposed Architecture (2 stream Fusion):
Predicion Samples:
