Dynamic networks, characterized by their continuously changing nodes and edges, play a vital role in various real-world applications. Community detection in dynamic networks, also known as dynamic community detection, is an active research topic. However, most existing methods still face two key problems. Firstly, community membership evolves over time, but few methods effectively exploit this dynamic information. Secondly, they completely rely on the assumption of temporal smoothness, often overlooking the abrupt changes of local topology, referred to as local non-smoothness. In view of this, we propose a contrastive learning based model DCDMT for dynamic community detection. Specifically, for the first problem, we introduce the detection mechanism of community member transition, which can distinguish stable and unstable community members between snapshots. By treating these nodes as the corresponding positive and negative samples, we develop the contrastive learning framework to optimize the dynamic community detection model. For the second problem, DCDMT uses the Hilbert-Schmidt Independence Criterion to improve node independence over time. This strategy helps capture local non-smooth features of nodes across snapshots. Experimental results on real-world datasets demonstrate that DCDMT outperforms state-of-the-art methods.