Hierarchical text classification (HTC) is a challenging sub-task of multi-label text classification due to its complex label hierarchical structure. Existing methods primarily focus on better utilizing hierarchical information, while overlooking a major issue, i.e., data imbalance, particularly the scarcity of training data for lower-level in the hierarchical structure, which results in poor classification performance. This paper proposes a topic-driven data augmentation method to enrich the training data for lower-level labels. Our approach is inspired by the topic detection and tracking (TDT) task, which involves finding and following new events in continuous data streams. Specifically, for each training sample of lower-level, we mine topic-consistent, semantically relevant texts from external knowledge bases and incorporate these texts as new training instances into the dataset. Additionally, we construct a dual-stream BERT-based top-down hierarchical classification framework to better model label hierarchical dependencies. Extensive experiments on three benchmark datasets show that our approach achieves state-of-the-art (SOTA) performance on the Macro-F1 metric. Our code and data are available online at https://github.com/oohs006/HierTDT