In this paper, we propose a new approach to multi-people activity recognition in outdoor scenes. The proposed method is based on Hidden Markov Models with parameters of reduced dimensionality. Most existing work is based on HMMs and DBNs, and focuses on the interactions between two objects. However, longer feature vectors of HMMs usually lead to covariance matrix singularity in parameter learning and activity recognition. Moreover, arbitrary structure of DBNs can introduce large computational complexity. Compared with former works, the proposed method named PCA-HMMs reduces the dimensionality of the model parameters while retains most of the original variability, and thus avoids overflowing and weakens the constraints on observations in conventional HMMs. The experimental results proved that the modified HMMs are effective solutions for multi-people interactive activity recognition.