Visual SLAM systems face significant challenges under fast camera motions, where limited inter-frame overlap leads to tracking failures and reduced accuracy. To address this, we propose a detector-free feature-matching SLAM framework tailored for rapid movements. Our approach introduces (1) a keyframe graph module that enforces matching consistency by linking keyframes through temporal and spatial constraints, followed by joint optimization, and (2) a semi-dense matching module leveraging detector-free relative encoding to establish robust correspondences under low overlap, with RANSAC filtering outliers. The system integrates keyframe selection and bundle adjustment for pose and map optimization. Experiments on public benchmarks demonstrate improved robustness in both RGB and RGB-D modes compared to state-of-the-art methods, effectively mitigating initialization failures and tracking loss in high-dynamic scenarios.