The Visual-inertial odometry (VIO) system is one of the core technologies in autonomous driving systems, it can provide pose estimation and motion trajectory of mobile robots by a monocular camera and a low-cost inertial measurement unit (IMU). However visual-inertial odometry usually extracts point features from images by visual descriptors such as ORB, FAST or Shi-Tomasi, while these methods rely on high-texture scenes which is not commonly appear in indoor scenarios. This paper provides a low texture indoor environment data, in which VIO system performs poorly by hand-crafted descriptor features in the front-end; then we use different visual features including line features and deep learning features from images and evaluate their performance in our data. Through the experiments we did, we propose a high-accuracy VIO system based on a combination of different visual features. Extensive experiments have been done on more public datasets, which demonstrate our VIO system remains robust and accurate performance in various scenarios, with only a monocular camera and low-cost IMU unit.