Recognizing Fall Actions from Videos Using Reconstruction Error of Variational Autoencoder

In this paper, we propose a method for detecting fall actions using a variational auto-encoder (VAE) with 3D-convolutional residual blocks. The VAE learns a distribution of Activity of Daily Life (ADL) data, and the reconstruction error is used to detect fall actions. The proposed method is a kind of unsupervised learning method with weakly labeled data for solving the problem of imbalance between the amount of fall data and that of ADL data. Furthermore, we propose extracting a human region from an entire image using skeleton information and aligning motions using the same joint point so that the neural network can focus on learning human motions, which can enhance the accuracy of fall detection. The results of experiments showed that our method achieved a competitive level of accuracy and better generalization ability compared to supervised learning with well-labeled data.