Physical activity is widely known to be a key factor towards achieving a healthy life and reducing the chance of developing certain diseases. However, there are many different physical activities having different effort requirements or having different benefits on health. The reason why automatic recognition of physical activity is useful is twofold: first, it raises personal awareness about the physical activity a user is carrying out and its impact on health, allowing some apps to give proper credit for it; second, it allows medical staff to monitor the activity levels of patients. In this paper, we follow a proven activity recognition chain to learn a classifier for physical activity recognition, which is trained using data from PAMAP2, a dataset publicly available in UCI ML repository. Once a machine learning dataset is created after signal preprocessing, segmentation and feature extraction, we will explore and compare different feature selection techniques using genetic algorithms in order to maximize the accuracy and reduce the number of dimensions. This reduction improves classification times and reduces costs and energy consumption of sensor devices. By doing so, we have reduced dimensions to almost a half and we have outperformed the best results found so far in literature with an average accuracy of 97.45%.