Speakers
Description
In this study, the method that recognizes human actions is analysed. It uses manually created shape features, and is tested in combination with a neural network-based classifier. We assume an application scenario involving the recognition of physical exercises as one of the preventive measures to reduce the risk of non-communicable diseases in the elderly. A popular action recognition dataset is used, as it contains activities corresponding to selected exercises. In addition to the application, the paper focuses on the study of combining neural network classifier with manually created shape features extracted from a small dataset. The main steps of the approach include calculating shape descriptors for all moving foreground objects extracted from video frames, then using these descriptors to construct feature vectors, and ultimately applying the Fourier transform to create representations of action sequences. A coarse classification step is included, which distinguishes between actions performed in place and actions with changing object's locations. The final classification is carried out using a neural network and a leave-one-actor-out cross-validation procedure. This paper presents experimental results on classification using the proposed approach based on simple shape descriptors and a feed-forward neural network with a single hidden layer. The averaged accuracy exceeds 97%.
