Video-based person re-identification （Re-ID） is a widely used task in computer vision. At present， most video-based Re-ID methods are based on supervised learning， which requires intensive manual annotation， and is very expensive and not suitable for real-life scenarios. In this work， an unsupervised video-based person Re-ID method based on diversity constraint and dispersion hierarchical clustering is proposed. First， each sample is regarded as a single cluster， and both the inter and the intra-cluster dispersions are combined to perform bottom-up hierarchical clustering. Second， clusters with small inter-cluster and intra-cluster dispersions can be prioritized for merging. At the same time， the diversity constraint is added to the clustering criterion to balance the number of samples in each cluster. Finally， the model is dynamically updated by using a linear feature memory. Experimental results on two public benchmark datasets， including Mars and DukeMTMC-VideoReID， show that compared with the state-of-the-art unsupervised video-based person Re-ID methods， the proposed method has some improvement in performance.