Human activity recognition (HAR) is an important research issue for pervasive computing that aims to identify human activities. Extracted features from raw sensors are often large and some of them can be irrelevant and redundant. Therefore, it's important to perform feature selection to select the most relevant features in order to increase the recognition accuracy. However, classical feature selection methods are generally linear and sequential and they do not consider existing dependencies and interactions among activities (classes). To overcome this shortcoming, a feature selection based on Choquet integral for HAR is proposed in this paper. The Choquet integral is a non linear and a non additive method. It's employed to determine scores for features by modeling interactions between activities through the fuzzy measure theory. Classification results on HAR dataset using Random Forest classifier indicate that the recognition accuracy remains stable using half of the features. Moreover, classification performance is further improved