Visual Odometry (VO) estimation is a component study of Visual SLAM in robotics and building applications to support visually impaired people in finding, tracking, and navigating. Deep learning with its outstanding advantages in accuracy has been widely applied to solve VO estimation problems by solving computer vision problems. However, it requires a huge amount of data to learn about the diversity of different environments. We make two main contributions: (1) The first is to publish the TQU-SLAM Benchmark dataset (TQU-D) collected from the hallway of three connected buildings on the second floor of TQU, Vietnam. These three buildings are interconnected with a movement length of 230m. This dataset includes 160,631 RGB-D frame pairs captured with 8 movement times (4 DI times and 4 VE times), and published data includes RGB-D data and 3D ground-truth (GT) camera motion trajectory/pose. The 3D original data is built based on manually prepared data on 2D images and the point cloud data construction method. (2) The second is to perform fine-tuning of the VO estimation model based on the end-to-end MLF-VO framework with backbones such as Resnet-18, and Resnet-34 on 12 cross-divided subsets of the TQU-D, and the estimated error is from 16.97 to 57.61 m. VO estimation results are evaluated and presented in the available. With these results, the database needs to be standardized and improved in the future.