The aim of this study is to test the effect of binary shape generalisation on the effectiveness of human action classification, with a focus on physical exercises recommended for the elderly to maintain physical fitness. An action recognition method based on shape descriptors and the Fourier transform is applied. The input data is represented in the form of a set of binary images, each containing a single shape that is either a detailed silhouette or a generalisation of it: a convex hull, an axis-aligned minimum bounding rectangle or an arbitrarily oriented minimum bounding rectangle. The individual shapes corresponding to each video sequence are represented numerically and the resulting representations are combined into a feature vector. The transformed feature vectors are then subjected to classification applying the nearest neighbour algorithm. The analysis of the experimental results makes it possible to determine the degree of relevance of simple shape features in an action recognition scenario and their applicability in a real-time system.