Robot Localization Based On Computer Vision

Automatic localization, which is of great value in research and application, is the core issue of robot navigation. Traditional localization methods based on global positioning system or dead reckoning have some limitations. Vision based localization method can remedy those defects.This paper studied robot automatic localization methods based on image features, compared some different kinds of features which could be well used in robot localization and then proposed two methods to improve the existing visual odometry algorithm.The first improvement is a precise self-localization algorithm based on improved Harris feature. We presented the improved strategy of extracting Harris corners in Gaussian scale spaces. Using the IMU prior data to estimate the areas of the tracked features ensures the accuracy of feature-tracking. Finally, quaternion and least square method were used to estimate the attitude information of the vehicle. Results of the experiments show that the algorithm fuses IMU data and visual information excellently, makes feature matching and tracking stably, avoids the drift problem of IMU data and achieves high localization accuracy.The other improvement integrates the advantages of Harris corner and SIFT feature. A fast and robust image feature called HSIF and its descriptor were proposed. This algorithm improves the features'adaptability to scale change and reduce the disadvantages of the Harris region-based matching. Meanwhile, we used a series of error analysis parameters to remedy the simple analysis in traditional methods. During data analysis, we compared our localization results with the ground truth, and our results showed the good efficiency and stability. Comparing with existing research results of robot localization, this algorithm demonstrates the high accuracy and practicality.