| The location and recognition of feature points is an important link in camera calibration and PNP(Perspective N Point)position solution,and its recognition and location results directly affect the whole visual measurement effect.In practical application,the recognition accuracy and positioning accuracy can be ensured by extracting feature points of cooperative coded targets.At the same time,the catadioptric omnidirectional camera can obtain a horizontal field of view of 360 at one time with a simple structure,which is gradually highlighted in the visual measurement of large field of view.Therefore,this paper studied the location and recognition method of coded targets in catadioptric images.In this paper,the deep learning convolutional neural network is introduced into the recognition of coded targets in catadioptric images,and the Mask R-CNN model is used to realize the two tasks of recognition and segmentation simultaneously.It can not only identified the identity information of coded targets,but also obtain the pixel edge of coded targets.At the same time,in order to solve the problem of insufficient samples in the training process of Mask R-CNN model,a method of generating simulated catadioptric images as training data sets based on spherical unified model was proposed.Experimental results show that,the recognition accuracy of coded targets in catadioptric images based on Mask R-CNN reaches 97%,the recall rate reaches 96%,and the average processing time of a single image is 0.2s The detection accuracy and robustness are better than those of traditional coded targets recognition methods.In order to further improve the positioning accuracy of coded targets,combined with catadioptric camera projection model,an iterative optimization method of curve equation and edge sub-pixel points is proposed,which improves the edge extraction accuracy of catadioptric images and obtains high-precision sub-pixel coordinates of feature points.Experimental results show that,this method is robust to fitting edge noise and camera position changes,and the sub-pixel extraction accuracy is 0.3 pixel.Aiming at the problems of large distortion of coded targets in catadioptric images and low recognition and positioning accuracy of traditional methods.In this paper,the Mask RCNN model is introduced,and the model parameters are successfully trained by generating simulation samples.According to the characteristics of straight line projection in catadioptric reflection system,the sub-pixel coordinates of feature points are obtained by iterative optimization,which provides research ideas for the application of coded targets in catadioptric vision measurement and navigation,and has strong engineering application value. |
