A Fall Detection System Based on Convolutional Neural Networks

The paper proposes a system to detect the accidental falls based on Convolutional Neural Networks (CNN). The original video is captured using a normal digital camera. For the preprocessing, the video is split into several consecutive frames and then subtract the background from each frame based on K-Nearest Neighbor (KNN) algorithm to extract the moving object. Some noise is reduced by eroding and dilating. The system chooses the VGG-16 network which is one of the models of CNN for action recognition. The preprocessed frames are used as input to the networks followed by a novel three-steps training phase. First, the transfer learning is applied to do the full training in the ImageNet dataset to learn a generic feature extractor. Then the fine tuning is implemented by retraining the network in the UCF101 dataset to learn features to represent human motions. Finally, the network is fine-tuned by retraining in the UR Fall Detection (URFD) dataset which contains 'fall' and 'not fall' features. After that, the system focuses on the binary problem of fall detection. The generality test and some evaluations of the system are stated at the end of paper.