In industry, ergonomic assessment is currently performed manually based on the identification of postures and actions by experts. We aim at proposing a system for automatic ergonomic assessment based on activity recognition. In this paper, we define a taxonomy of activities, composed of four levels, compatible with items evaluated in standard ergonomic worksheets. The proposed taxonomy is applied to learn activity recognition models based on Hidden Markov Models. We also identify dedicated sets of features to be used as input of the recognition models so as to maximize the recognition performance for each level of our taxonomy. We compare three feature selection methods to obtain these subsets. Data from 13 participants performing a series of tasks mimicking industrial tasks are collected to train and test the recognition module. Results show that the selected subsets allow us to successfully infer ergonomically relevant postures and actions.