Fall Detection Based on RetinaNet and MobileNet Convolutional Neural Networks

Abstract

The problem of falling is a major health problem resulting in serious injuries and sometimes lead to death especially for elderly. Elderly people aged over than 75 are exposed to accidental deaths due to falls. Approaches based on computer vision give a promising and an effective solution for detection human falls. This paper presented a method for fall detection which based on combining convolutional neural networks RetinaNet and Mobilenet in addition to handcrafted features. Traditional human detection methods may result in human shape deformation which affect the performance of fall detection frameworks. Therefore, the proposed framework depends on RetinaNet for detecting humans with shorter computing time and higher accuracy compared with the traditional human detection methods. Then, the proposed framework relies on handcrafted features to represent shape and motion properties of the detected human. The proposed framework extracts aspect ratio and head position as shape features and motion history image as a motion feature of the detected human to create the feature map. This feature map is used in training MobileNet network to classify the human motion into fall or not-fall. The proposed framework is evaluated using UR and FDD datasets and the experimental results proved the efficiency of the proposed framework achieving up to 98% accuracy compared with the state-of-the-art methods.