Abstract: Aiming at the problem of high-precision positioning of mobile robot in Global Navigation Satellite System GNSS denied environment, An adaptive thresholding adaptive and generic accelerated segment test AGAST feature detection algorithm is proposed to improve the visual front-end of a vision/inertial guidance fusion localization system for mobile robots. The algorithm improves the visual odometry computation method by local histogram equalization and adaptive threshold detection, improves the quality of feature point extraction, and enhances the positioning accuracy and stability of visual odometry in complex environments. Visual odometry and inertial navigation system are fused based on factor graph optimization algorithm to realize high-precision positioning of mobile robot. The results show that, compared with the mainstream VINS-Mono algorithm, the proposed algorithm improves the positioning accuracy by 22.8% in the experiment of indoor data set and 59.7% in the experiment of outdoor data set, the proposed algorithm perform better than VINS-Mono algorithm in both two experiments and it can provide better positioning services for mobile robots.