The development of biomedical engineering causes increase of the quality of medical services, and one of the areas of biomedical engineering in which technology development has a significant impact is researching different aspects of physical activity: daily physical activity, exercise, diagnosis, therapy and rehabilitation. Basic motivation for physical activity monitoring is use in sports, but it is also often applied to the recognition of older people's activity, their balance, with the aim of providing preventative procedures in case of loss of balance. For example, for accurate fall detection it is also necessary to identify which activities were preceding. The main aim of this thesis is to create an algorithm for classification of physical activities by computing features from acceleration sensory node measurements with sensory node set on the chin while performing different physical activities. For this reason, the acceleration signals of nineteen respondents – volunteers were recorded. Each person has performed thirteen static or dynamic physical activities: standing, walking, fast walking, running, fast running, walking down the stairs, walking up the stairs, high and low jump, sitting, getting up from the sitting position, lying down and rising from the lying position. The Shimmer3 sensor node for acceleration measurements in x, y and z directions was used, and the algorithm was created in the MATLAB programming environment. As a result, necessary methods of data processing have been defined since it is left free to set the position of the Shimmer sensor node on the chin. It is assumed that everyday users will rarely set the sensor on the right position, so the aim is to calibrate the data in relation to the position in the transfer plane. In addition to this, the features that differentiate each activity in the best way are defined, with the selection of features having a significant impact on the classification. Algorithms that provide the best results for this application are also defined. Regardless of the features and classification methods, it is easiest to classify actions that differ in dynamics such as standing, jumping, and lying down. Measuring acceleration values from the sensor node, their processing and machine learning can provide a high quality classification of physical activity. The classification algorithm for identifying physical activities of elderly people with aim of falls prevention has been proposed and tested.