Knowledge-grounded dialogue methods based on small language models (SMLs) hold significant research value, yet face multiple challenges including limited generalization and reasoning capabilities compared to LLMs, model alignment difficulties, and frequent low-resource application scenarios. To address these challenges, this paper presents a novel three-stage knowledge-enhanced and latent-topic learning framework for low-resource dialogue generation with SMLs. The SMLs-based dialogue model initially is pre-trained on knowledge-grounded dialogue datasets, and then a supervised fine-tuning for the model is performed on low-resource dialogue datasets. Model domain adaptation issues are alleviated through application of a noise-injection-based denoising pre-training approach to dialog texts. A retrieval-augmented knowledge generation module is introduced to produce context-relevant knowledge, effectively mitigating knowledge scarcity in low-resource settings. To solve topic drift and intent deviation, a method via conditional variational autoencoder-based latent topic learning is proposed to guide response generation. To improve model performance, self-critical sequence training for reinforcement learning is employed for the dialogue model on low-resource datasets. Experimental results demonstrate that the proposed method significantly outperforms baseline approaches across multiple metrics evaluating response generation quality, including diversity, coherence, and accuracy.