Deep Learning-based Vision System Modeling Research With Calibration Technology

When autonomous vehicles are driving on the road,if they cannot accurately collect important external information such as pedestrians,vehicles,and puddles,it is highly likely to cause property damage and casualties.In order to solve the accuracy problem of information collection,this article studies the combination of deep learning and visual imaging models to obtain accurate object positions,thereby improving the reliable driving of autonomous vehicles.The existing camera measurement methods have good results,but there are still the following problems:in the process of camera calibration,it is necessary to consider the physical characteristics of the camera and the impact of external environmental information on camera calibration.Generally,the method of combining multiple linear processes to solve the internal and external parameters of the camera has problems such as more intermediate solving processes and reduced measurement accuracy.A high-precision,robust measurement model and calibration technique utilizing deep learning mapping relationships have been designed to address the above issues.According to the design requirements,the mapping model principle of traditional cameras was analyzed,the correspondence between pixel points and world coordinate points was studied,and the advantages and characteristics of deep learning were analyzed.A convolutional neural network visual measurement model that meets the data features was constructed;Adopting filtering,morphological processing,and corner extraction methods to collect pixel coordinates,measuring 3D world coordinates based on precision instruments,and then integrating the data;In order to improve the camera calibration accuracy,normal distribution and polynomial fitting methods are introduced to analyze the data comprehensiveness and accuracy;In order to get the parameters of the convolutional neural network model efficiently and quickly,and process the data,polynomial regression method is used to analyze the data comprehensiveness and error,support vector regression method is selected to process outlier,one by one troubleshooting method is used to eliminate duplicate data,and the maximum and minimum method is selected to normalize the data.The normalized data is proportionally allocated into a training set and a testing set.The training set data is used to solve the parameter values of the convolutional neural network visual model,and the testing set verifies the model.In this paper,the calibration platform of dual port vision for 3D environment measurement system is built,and the improved measurement imaging model is calibrated and data validation experiments are carried out.The experimental results show that the direct calibration method using deep learning visual imaging models in this article can simplify the calibration process to a certain extent.Compared with traditional measurement methods,the accuracy in theX_w、_wY andZ_w directions is reduced by 12.852mm,meeting the measurement requirements of autonomous driving and advanced assisted driving systems for external environmental information.