Visual guided deep learning scheme for fall detection

Fall detection is an important problem in the field of public health care, which is especially crucial for instant medical service delivery to the injured elderly due to falls. Ambient camera based fall detection has been a recognized non-intrusive and publicly acceptable method, where video data is employed to discriminate fall event from daily activities. Fall detection with videos usually requires a large dataset to extract features and train the classifier. However, it is hard to collect free-living environment fall data and instead simulated falls by young people have been collected to construct the training dataset, which is controlled intentional behavior and restricted to limited quantity of samples. In addition, the existing video based fall detection methods need segment the subject first, which is inclined to be influenced by image noise, illumination variation and occlusion. To address these problems, a three dimensional convolutional neural network (3D CNN) based method for fall detection is developed which only uses kinetic data to train an automatic feature extractor. Besides the spatial feature in 2D image, the motion information from the video could also be encoded by the three dimensional convolutions over the frames. A LSTM based spatial visual attention scheme is then incorporated, which could enable the network to focus on the key regions. Sports dataset Sports-1M with no fall examples is employed to train the 3D CNN and the visual attention model is trained on the small Multiple Cameras Fall Dataset. Then the visual attention based 3D CNN is employed to extract the features from the videos with fall event and implement fall detection. Experiments have shown the superior performance of the proposed scheme on fall dataset with high detection accuracy of 100%.