Unreliable feature extraction and matching in handcrafted features undermine the performance of visual SLAM in complex real-world scenarios. While learned local features, leveraging CNNs, demonstrate proficiency in capturing high-level information and excel in matching benchmarks, they encounter challenges in continuous motion scenes, resulting in poor generalization and impacting loop detection accuracy. To address these issues, we present DK-SLAM, a monocular visual SLAM system with adaptive deep local features. MAML optimizes the training of these features, and we introduce a coarse-to-fine feature tracking approach. Initially, a direct method approximates the relative pose between consecutive frames, followed by a feature matching method for refined pose estimation. To counter cumulative positioning errors, a novel online learning binary feature-based online loop closure module identifies loop nodes within a sequence. Experimental results underscore DK-SLAM's efficacy, outperforms representative SLAM solutions, such as ORB-SLAM3 on publicly available datasets.