An improved graph optimization-based mapping algorithm was proposed to address the limitations of traditional graph optimization methods, including poor environmental perception, indistinct feature extraction, and large cumulative errors. In this approach, a LiDAR and a depth camera was first calibrated, and their data was processed and fused to preliminarily reduce sensor errors and data redundancy. A distance evaluation function was then incorporated into principal component analysis (PCA), and an adaptive threshold was set to select point cloud feature points. The initial pose transformation matrix was obtained using IMU pre-integration for preliminary scan matching, after which a point-to-plane iterative closest point (ICP) algorithm was applied to derive the final pose transformation matrix. Loop closure detection was performed using Monte Carlo localization (MCL) combined with odometry data to eliminate cumulative errors. Experiments were conducted using a smart wheelchair in real-world scenarios. The results indicate that the proposed algorithm exhibits superior environmental perception compared to traditional graph optimization methods and effectively reduces cumulative errors in large-scale environments, demonstrating noticeable improvements over conventional approaches.