The wide use of wearable devices such as smartwatches and fitness trackers enables gathering of various information about the wearer's physical activities. This information, in turn, can produce valuable insights, such as recognizing activities being performed and interacting with smart environments based on recognized activities. In order to recognize the activity being performed, a machine learning (ML) model can be built. However, this process can be complicated since there are many different ML algorithms that have various hyperparameters that impact the algorithm's performance, and for different devices in each condition of their usage, a different machine learning algorithm may be more suitable. The task of simplifying a machine learning model construction is in the focus of the automated machine learning (AutoML). However, the methods used in AutoML are in turn very time- and resource-consuming due to the large search space formed by huge number of possible algorithms and their hyperparameters. In this work, we suggest a method for building machine learning models for physical activity recognition based on the AutoML approach enhanced with meta-learning and hierarchical grouping of algorithms. The meta learning technique forms the hierarchy and determines the appropriate group of algorithms for each case of human activity data. The selected group is then used to limit the search space and thus reduce the resources required for the search. The experiments performed on the AutoML library Auto-Sklearn have shown that limiting the search space with a group of algorithms enables the reduction of the time required for model building and that the use of upper-level groups of algorithms leads to an increase in search time compared to the groups of the bottom level.