Abstract: A visual-inertial SLAM (simultaneous localization and mapping) optimization algorithm is proposed to address the problem that visual SLAM is not robust and accurate for indoor robots in weak texture environment. Firstly, the deep learning module based on the attention mechanism is used to directly match features between two adjacent frames, and then the traditional feature detection methods are adopted to solve the problem that enough feature points can't be extracted. Secondly, a camera depth confidence model is established to reduce the drift error of long-distance features by introducing the depth confidence probability of the spatial points before calculating the inter-frame pose transformation of the camera. Finally, the deep confidence of all matched feature points involved in back-end optimization is used as the piecewise threshold of dynamic robust kernel function to optimize the traditional bundle adjustment and coordinate the overall motion trajectory. It is shown in real scene experiments that the proposed algorithm is significantly robust in weak texture environments. Compared with VINS-RGBD algorithm, the absolute trajectory error of the proposed algorithm is reduced by 50.38% and the relative trajectory error is reduced by 85.75%.