Conventional visual-inertial SLAM algorithms often suffer from poor feature quality in dim and low-texture environments, leading to unstable tracking, frequent tracking loss, and low localization accuracy. To address this, this paper proposes a stereo visual-inertial SLAM algorithm for UAVs that integrates texture awareness and adaptive feature tracking. First, a noise suppression factor is introduced to prevent noise amplification during image enhancement, and a texture-aware control weight is designed to adaptively adjust the enhancement strength based on local texture richness, applying stronger enhancement in texture-rich regions while keeping the enhancement low in texture-sparse areas, thereby improving overall image contrast while effectively suppressing noise and preserving useful details. Second, the XFeat network is used to replace hand-crafted features, and a density-feedback threshold adjustment mechanism is introduced to address uneven feature distribution by adaptively adjusting the detection threshold according to local feature density, enabling more features to be detected in sparse regions while maintaining a higher threshold in dense regions to retain high-quality features, thereby improving the stability and uniformity of feature tracking. Finally, these improvements are integrated with the VINS back-end optimization to build a complete SLAM system. Evaluations on the EuRoc dataset show our method achieves 29% and 13% higher accuracy than VINS-Fusion and SuperVINS, respectively, in challenging dim and low-texture scenarios. Real-world experiments demonstrate a closure error of 0.486 meters, confirming its superior performance.