Abstract: To address the significant challenges posed by low-texture and dynamic scenes in building environments for robotic localization, a multi-source data fusion based localization method tailored for structured building scenarios is proposed. By judiciously incorporating multi-primitive visual features, Manhattan World Hypothesis constraints, and IMU (inertial measurement unit) pre-integration constraints, the proposed localization system effectively tackles issues such as sparse features and dynamic personnel interference encountered by intelligent bed-and-chair robots during operation. Overall, a parallel strategy combining high-frequency front-end pose estimation and low-frequency backend optimization is adopted to enhance system real-time performance. The front-end odometry provides relatively accurate pose estimates within short periods, which are further refined by backend optimization using the Manhattan World Hypothesis. Additionally, a multi-camera joint localization strategy is introduced innovatively to address the limitations of the Manhattan World Hypothesis in pose estimation applications. Two cameras with different orientations are utilized, one to capture rich texture features on the ground and the other to identify planes in the upper vertical space that satisfy the Manhattan World Hypothesis, which mitigates the impact of non-Manhattan objects on localization. Experimental results demonstrate that, compared to the existing methods, the proposed method significantly improves the localization accuracy and robustness while ensuring real-time performance, providing robust support for the application of intelligent mobile bed-and-chair robots in building environments.