| Accurately estimating position and posture of Intelligent vehicle is an important prerequisite for autonomous.Currently,traditional visual odometry and integrated positioning scheme(Global Navigation Satel-lite System-Inertial Navigation System,GNSS-INS)have met the basic needs of navigation and positioning technology.However,the traditional visual odometry is easy to accumulate errors in the scene of changing light intensity.GNSS positioning is often confused by weak signal strength and high signal noise.Semantic features have good performance in constraints and robustness.Therefore,it is important to study the visual position estimation method combining GPS and semantic information.In order to establish robust semantic constraints,a semantic observation model is designed based on semantic category characteristics.An ERFNet semantic segmentation network is built,and trained with warmup-poly learning rate adjustment strategy.Then experiments verify the effect of the model.We build observation model of semantic feature based on local neighbor distance which is calculated by pruning optimization breadth-first search,and pixel offset of semantic segmentation.The multiple weights representation method and weight increment update method of map point semantic category are designed.We establish the calculation method of category weighted semantic error and deduce jacobian matrix of semantic error term to Lie algebra,and the validity of the modeling method is verified through experiments.To improve the constraint efficiency of semantic feature association,a semantic visual odometry based on mid-term sliding window is designed.Considering the insufficient semantic constraints,a visual odometry combining semantic features with ORB feature points is designed.A quadtree structure is used to extract FAST key points.KNN matching algorithm based on ORB feature or fast matching algorithm based on word bag model is used to achieve correct matching of feature points for different scenarios.A step-by-step optimization structure of initial position and posture optimization and local optimization is used.We propose a medium sliding window structure to establish the semantic feature associations.A graph optimization model of local optimization for joint semantic information and ORB features is constructed,and it is solved piecewise.Finally,experiments show that the mid-term sliding window structure can establish effective semantic feature association.To reduce the complexity of GPS information fusion algorithm,a GPS fusion algorithm based on displacement error is presented.According to the positioning requirements,the differential GPS positioning is selected to obtain GPS data,and the GPS error items are established through the position deviation.In order to simulate the GPS signal missing scene,local optimization process combines GPS data of part of the time to form a tightly coupled pose estimation problem with multi-information fusion which is solved by graph optimization method.The results show that the positioning accuracy of this method is better than that of semantic visual odometry. |
