Research On The Visual Pose Estimation Method Of Intelligent Vehicles In Urban Environment

Accurate and real-time pose estimation is a necessary prerequisite for autonomous navigation of intelligent vehicles.At present,the integrated positioning scheme(GNSS-INS)of global navigation system and integrated navigation system and and high-definition map(HD Map)have been able to meet the positioning requirements of intelligent vehicles.However,in the face of complex urban environments,GNSS-based positioning is unreliable due to factors such as signal occlusion and multipath effects.At the same time,the cost of making and updating commercial HD map is high.In view of the superior performance of vision method in robot and UAV pose estimation.Therefore,it is of great significance to carry out research on the intelligent vehicle visual pose estimation method in the urban environment.In order to construct a robust visual pose estimation model,a visual odometry based on feature points was systematically designed by combining the vehicle motion characteristics.Mathematical modeling of camera imaging and distortion was carried out,using a threshold self-adjusting FAST key point detection method to adapt to different quality images,nonmaximum suppression was used to solve the problem of corner aggregation,the Hessian matrix of the candidate points was calculated,and the edge points were eliminated according to the corresponding relationship between the eigenvalues of the Hessian matrix and the edge points of the image.The two-way matching strategy was used for feature point association,the Ackerman steering vehicle motion model and motion characteristics were integrated into the pose estimation,the basic model of intelligent vehicle pose estimation was obtained through system integration,and the validity of the model was verified by experiments.In order to improve the positioning accuracy of the system in the urban environment,a multi-information fusion method of GPS and target detection assisted vision was proposed.A graph optimization model of visual odometry and GPS information fusion was constructed.The residual items in the model were calculated to build a nonlinear objective function.According to the accuracy of GPS,different weight factors of GPS information were assigned,the vehicle pose after fusion is calculated by minimizing the objective function.The detection algorithm based on deep learning was used to detect dynamic objects in the environment,and the dynamic feature points were removed before the matching of feature points to reduce the impact of dynamic objects on the positioning accuracy.A method for distinguishing "pseudo-dynamic objects" is proposed to solve the problem that the target detection algorithm regards stationary vehicles and pedestrians in the environment as dynamic objects.The experimental results show that this method can overcome the disturbance of dynamic objects and has high positioning accuracy in urban environment.In order to improve the computational efficiency of the system,parallel computing was designed for tracing thread.The time ratio of each part of the tracking thread was analyzed to determine the feature point extraction,matching and positioning as the modules to be optimized.A parallel computing framework for feature point extraction and matching was established on CPU-GPU heterogeneous platform to achieve the acceleration of this part.For the positioning part,a 3D interior point detection strategy was proposed to realize the parallel search of map points,and the saturated linear kernel function was used to act on the reprojection error,so as to reduce the influence of abnormal points on the positioning accuracy and realize the parallel calculation of pose optimization.The experimental results show that the average acceleration ratio of feature point extraction and matching is 6.5 times,and the overall computational efficiency of the system is about 7 times higher than that before acceleration.