Research On Construction Technology Of Test Scene Sample Library For Unmanned Vehicle Vision System

The unmanned vehicle industry has become the focus of attention from all over the world.And it is listed as a national key development direction by "China Intelligent Manufacturing2025".In recent years,there have been a lot of achievements in the development of unmanned vehicles,such as: express unmanned vehicles,sweeping unmanned vehicles,unmanned buses,and unmanned private vehicles.The Tesla and Uber self-driving vehicle accidents illustrate the need for closed testing of unmanned vehicles.The vision system is one of the most important perception methods for unmanned vehicles to obtain the external environment.Therefore,the testing of the visual performance of unmanned vehicles is an important part of unmanned vehicle testing,and it is an important indicator to ensure the safety of unmanned vehicles and evaluate the substitution of unmanned vehicles into the market.Based on the research of the virtual scene construction method,principle and algorithm of the unmanned vehicle vision test,this paper designs and proposes an unmanned vehicle test scheme through combining virtual and real.The thesis takes the method in the construction of the scene sample library of the unmanned vehicle visual virtual test as the starting point,and studies the key technologies involved in sample collection,cleaning,feature recognition and simulation modeling.It has more important theoretical significance and practical value for the visual acceleration test of unmanned vehicles.The main work of this paper includes the following parts:Firstly,the overall block diagram of the construction of the unmanned vehicle visual test scene sample library is proposed.Using analogy to human visual perception and testing,an overall model of unmanned vehicle visual system testing is proposed,and the key technologies of each part of the design are further analyzed.Secondly,in view of the problems of sample errors,duplication or similarity in the construction of the static vision test sample library,data cleaning rules are defined,and a traffic scene image deduplication algorithm based on perceptual hashing under superpixel segmentation is proposed.After experimental verification,the algorithm has an accuracy of98.55% and has high accuracy,robustness and stability,simultaneously.Thirdly,the static vision test of unmanned vehicles is summarized into two parts: the recognition of basic image features such as texture,color,and shape,and the recognition of traffic scenes and traffic participants.On the one hand,it analyzes the domestic and foreign traffic sign data sets,constructs the traffic sign data set N-CTSDB20 of this article,and selects the SSD model for detection and recognition,so as to realize the automatic recognition of traffic signs.On the other hand,it analyzes the basic parameters that affect the perception of the unmanned vehicle’s vision system,builds basic vision test tools,and provides a sample basis for unmanned vehicle vision testing.Then,it analyzes the composition of traffic scene elements,determines the function and structure of the simulation platform,uses Unity 3D and Blender modeling software to realize scene modeling,and interactively designs to realize a virtual scene simulation platform based on Unity 3D engine.Finally,the constructed scene sample library and simulation platform are evaluated separately to verify their effectiveness.This paper realizes the construction of a sample library of unmanned vehicle vision system test scenes combining virtual and real.After experimental verification and evaluation,it can basically complete the unmanned vehicle vision test requirements.