Positioning Method of Intelligent Vehicle Based on Visual Multi-Feature Primitive Fusion

The robustness of visual localization algorithms is relatively poor when only visual point features are applied. Since the line feature is not sensitive to beam transformation, one description and extraction methodology based on long line features is proposed, in case that line features with length variations are obvious. ORB-SLAM2 framework is employed to fuse the point and line features, in order to complete the pose estimation and local optimization algorithm. Within the loopback detection module of the localization, spatial consistency and multipath fusion based on multi-features are applied to improve the recall rate of loop detection. Besides, path integration with correct re-position and camera pose transformation with multi-segment paths are successfully achieved. The real vehicle tests show that the point-line fusion algorithm has better robustness than single point localization algorithm. Also, the multi-path fusion algorithm contains higher mileage rate during the public dataset tests. The check result reaches 99.91% which is 10.71% higher than the traditional algorithm.