An End-To-End unsupervised approach employing convolutional neural network autoencoders for human fall detection

In the past few years, several works describing systems for the prompt detection of falls have been presented in literature. Many of these systems address the problem of fall detection by using some handcrafted features extracted from the input signals. In the meantime interest in the use of feature learning and deep architectures has been increasing, thus reducing the required engineering effort and the need for prior knowledge. A fall detection method based on a Deep Convolutional Neural Network Autoencoder is presented in this work. This method is trained as a novelty detector through the end-to-end strategy. The classifier distinguishes normal sound events generated by common indoor human activity (i.e. footsteps and speech) and music background from novelty sound events produced by human falls. The performance of the algorithm has been assessed on a corpus of fall events created by the authors. Moreover a comparison was made with two different state-of-art algorithms both based on a One Class Support Vector Machine. The results showed an improvement on performance of about 11% on average.