We propose a topic model to better estimate activities from tweets. The whole estimation process consists of two phases: one is the cluster generation, and the other is the activity estimation. At the first phase, we obtain the expected trilayer clusters with the components: a topic layer, an activity layer and a word layer. Then, at the second phase, we utilize the activity-specific word distribution derived from the training results to learn the activities of testing tweets. To prove the feasibility of this model, we evaluate the precision of activity estimation using 35 activities to extract 23,988 tweets for training and 350 for testing, respectively. The experimental results demonstrate that the reasonable topic-specific activity distribution contributes to the cluster generation, and the proposed model exhibits the superiority in activity estimation.