Study On Natural Visual Landmark Aided Robot Localization

Automatic localization is the core issue of the robot research。Global Position System have high accuracy but greatly influenced by environment. Inertial Navigation System have advantages in self-localization, various kinds of navigation information. However, accumulated error will be its problem. Therefore, we firstly introduced the GPS/INS integrated navigation which is the most commonly used ways overcome the shortcomings of each system, while we applied the dual-GPS to Initialize in INS navigation. The experiments show that the system has high accuracy. However, in practical application, GPS signal will be blocked and the accuracy of the system will drop quickly.In this paper, we try to find the other ways of navigation to replace the GPS. On one hand, we present the global localization based on natural visual landmarks. We use natural scenes as visual landmarks, and match the images from the monocular camera with the landmarks. We build the image matching framework on the basis of GIST global features and SURF local features. Using the numbers of the features the images matched and their scale information as the rule to measure the similarity of images. We demonstrate the high accuracy and efficiency of the system. What’s more, we also show how the factors to influence the robust of the system.On the other hand, we build decoupled attitude estimation model to get much reliable attitude output, then integrated the wheel odometer with INS to reduce the accumulated error. Finally, we use Kalman Filter to combine natural visual landmarks localization and INS localization, while combining the Structure From Motion algorithm to correct the body attitude based Monocular vision. We use INS navigation as the main way and the visual localization just like the GPS data with high accuracy to correct the INS data.We present the experiment in typical outdoor grass environment and compared our localization results to the ground truth measured by total station. Our results demonstrate the high accuracy and practicality of the algorithm.