Loading...

Accuracy Improvement of Vision-Aided Gyroscope using Convolutional Neural Network

Shadravan, Shayan | 2017

630 Viewed
  1. Type of Document: M.Sc. Thesis
  2. Language: Farsi
  3. Document No: 50951 (19)
  4. University: Sharif University of Technology
  5. Department: Computer Engineering
  6. Advisor(s): Manzuri Shalmani, Mohammad Taghi
  7. Abstract:
  8. The growth of the knowledge of image processing and machine vision in recent years has led to many applications in various fields. One of the most important applications in machine vision is automotive navigation of vehicles and robots. The effective use of visual sensors to detect obstacles, routing, detecting the position of the robot, and mapping the environment is one of the most important goals in ground robotics. Few methods using sensors such as accelerometers, gyroscopes and global positioning systems, suffer from problems such as high costs, accumulative errors, dependencies on external systems, and the inability to be used in closed spaces. But with the use of the visual sensors, problems of cost, accumulative error, dependency and limitation on closed space are solved, and, if combined with other sensors, the accuracy of navigation in any moving device, whether robot or vehicle, in motion at all levels, will be increased effectively.
    In this study, we used the new methods of deep convolutional neural networks to increase the accuracy and speed of estimating the robot’s position. To evaluate the proposed method, the KITTI data set was used to compare the speed and accuracy with other existing methods. First, a pair of sequential image captured by a monocular camera is fed into the FlowNet2 network to extract optical flow images, and then the resulting images are given to the Visual Odometry network to calculate the robot's rotation angle and displacement. To enhance the processing speed of the neural networks, a high-speed GPU is used.
    The evaluation and comparison of the results showed that the accuracy of the rotation was very close to the state-of-the-art method and the translation accuracy increased up to 4%. Also, this method is up to 4 times faster that the previous method
  9. Keywords:
  10. Visual Sensors ; Convolutional Neural Network ; Rotation Estimator ; Visual Gyroscope

 Digital Object List

 Bookmark

No TOC