ABSTRACT Human physical activities play an essential role in many aspects of daily living and are inherently associated with the functional status and wellness of an individual, therefore, automatically and accurately detecting human activities with pervasive computing techniques has practical implications. Although existing accelerometer-based activity recognition models perform well in a variety of applications, most of them typically work by concatenating features of different domains and may fail to capture the multi-view relationships, resulting in degraded performance. To this end, we present a multi-view aggregation model to analyze the accelerometer data for human activity recognition. Specifically, we extract the time-domain and frequency-domain features from raw time-series sensor readings to obtain the multi-view data representations. Afterwards, we train a first-level model for each view and then unify the models with stacking ensemble into a meta-model. Finally, comparative experiments on three public datasets are conducted against other three activity recognition models. Results indicate the superiority of the proposed model over its competitors in terms of four evaluation metrics across different scenarios.