Abstract: Visual SLAM (simultaneous localization and mapping) algorithms are sensitive to illumination variation due to their dependence on the image brightness. To solve this problem, a semi-direct visual SLAM algorithm based on online photometric calibration is proposed. Firstly, an inter-frame pose estimation method based on photometric calibration is proposed, according to the principle of camera imaging. While solving the pose, photometric correction is performed on the original input image. Secondly, the most recent common-view keyframe is selected in the feature tracking process to improve the efficiency of feature point matching. Finally, the iterative optimization strategy in the back-end reprojection is improved to reduce the influence of illumination variation on the accuracy and robustness of SLAM. According to the experimental results on TUM and EuRoC datasets, the proposed algorithm outperforms LSD-SLAM and SVO 2.0 algorithms in terms of the trajectory estimation accuracy, especially for sequences of medium and difficult datasets. By comparing the trajectory estimation results of the proposed algorithm with the laser-based method in real environments, it is confirmed that the proposed method effectively improves the localization accuracy and robustness of traditional SLAM algorithms in uneven-lighting environments.