Research And Implementation Of Cone Recognition Technology Based On Monocular Vision

Based on the monocular camera,this paper conducts an in-depth research on the identification method of the cone that composes the driverless formula racing track.The traditional recognition technology and deep learning methods are used to construct the algorithm for the cone recognition,and the relative position information of the cone is calculated through the coordinate system conversion.The main work of the thesis is divided into the following three aspects:(1)To analyze the features of the cone,propose a cone recognition algorithm based on color feature and shape feature.Color extraction is performed through HSV color space and converted into a binary image,and then a series of desiccation processing is performed.The processed image is constructed based on the contour to construct a minimum circumscribed matrix,and the relationship between the two matrices divided by the white reflective strip is preliminarily identified.Afterwards,the slope function after the vertical projection is obtained by the least square method according to the symmetry feature,and the numerical analysis on both sides of the extreme point is used for identification.(2)The local coordinate system is established based on the camera,and the perspective transformation model is constructed through the basic principles of the camera coordinate system.The recognition cone is identified and the center point pixel coordinates are extracted.The perspective transformation matrix can be used to convert the center point pixel coordinates into the two-dimensional coordinate information of the cone relative to the camera.(3)Constructed a data set for identifying cones,and established a cone recognition model based on deep learning.Collect a large number of cone images in various environments,construct a data set for identifying the cone through manual annotation,and train the constructed model to realize the recognition of the cone.Based on the verification platform built by Songling Intelligent Chassis,this paper has carried out a real vehicle test.Experimental data shows that the recognition algorithm based on traditional image processing can achieve a recognition rate of more than 95% when adjusting the HSV parameter range and camera exposure time;the cone recognition model based on deep learning structure,the recognition rate is all under the same test standards More than 98%,the recognition rate is still more than 92% even when the two cones are improved;the positioning error of the cone is less than 10 cm,which can basically locate the position information of the cone.The research in this paper can effectively solve the problem of low recognition rate of cone recognition in different environments,and can meet the requirements of car visual recognition rate.