A New Approach For Human Fall Detection By Using A Convolutional Neural Network

Abstract

The prevalence of artificial intelligence techniques in the last decade, and the widespread of surveillance cameras, have contributed to the classification of human motions. Especially those that may have a negative impact on their health, and may, in worst cases, lead to their death. In this paper we have introduced a new method, called STTFI, that extracts a person's silhouette with spatial and temporal characteristics representing the motions and its quantities such as previous methods (MHI, BMI, ITMI). But it also represents the spatial and temporal characteristics of static poses that person passes during a fall. Then it has been used with a low-depth convolution neural network using learning transfer technique to classify the event as fall or not fall. The method gave results that outperformed most previous studies in accuracy, and some in sensitivity. the accuracy of this method is 99.02%, sensitivity 100%, and specificity 98.73%.