Research And Application Of Obstacle Detection Algorithm Intergrating Camera And Millimeter Wave Radar

In recent years,with the rapid development of computer science and multi-sensor fusion,self-driving car has become a research hotspot in the field of intelligent vehicles at home and abroad.Environmental awareness is equivalent to the eyes and ears of automatic driving,which is the basis of intelligent vehicle to achieve automatic driving,and also an important part of automatic driving.People’s expectations for self-driving car are getting higher and higher,so the function of autopilot system is becoming more and more complex,and the perception of the surrounding environment also requires higher requirements.At present,the existing research mainly focuses on "seeing" obstacles,and is only limited to the recognition of obstacles in simple scenes.In practical application scenarios,because the application scene is more complex,more interference factors and the impact caused by sensor noise is amplified,it is difficult to obtain ideal fusion results with the existing methods.Therefore,at present,the solution of automatic driving is to use the fusion of camera and millimeter wave radar to realize the complementary advantages of the two sensors,so as to achieve stable environment sensing output.However,if more than one sensor is introduced,the noise of the sensor will be introduced at the same time.The fusion system needs to have excellent performance in accuracy,stability and continuity,so as not to cause the error amplification of the subsequent system.In the existing fusion system.there is still challenges for further improvement in the research of the above aspects.In view of the above problems,this thesis combines the advantages of camera sensor in obstacle recognition and millimeter wave radar sensor in ranging.For the data fusion of the two sensors,based on the advantages of each sensor,a more accurate and stable obstacle detection effect is achieved.The fusion system proposed in this thesis firstly solves the problem of obstacle detection caused by different millimeter wave radars in unmanned driving effectively by clustering algorithm.Secondly,this thesis effectively solves the problem of wrong matching in the process of driverless driving by using the bipartite graph matching algorithm based on the idea of many to many matching and the three-dimensional physical world to image anti throw detection algorithm.Finally,simulation and real vehicle test verify that the above methods can improve the accuracy,stability and continuity of the fusion system.The main contents of this thesis are as follows:(1)In view of the mutual interference between forward millimeter wave radar and short-range angle radar in the forward detection area,which results in multiple fusion results of a real physical obstacle,this thesis proposes an improved clustering algorithm PB-DBSCAN algorithm based on DBSCAN.In this method,the requirements of output obstacles in the fusion system are considered comprehensively.By using the characteristics of two kinds of radar return obstacles and the characteristics of vehicle reflection points in automatic driving,the calculation method of neighborhood range is improved,and the transverse and longitudinal beam of density diffusion is reduced.The simulation and real vehicle test results also prove that the method can effectively and accurately deal with multiple radar signals The number is clustered to improve the accuracy of the output of the fusion system(2)In view of the problem that there are many target vehicles or large obstacles with multiple reflection points,obstacle information matching is prone to make mistakes.This thesis proposes a detection algorithm UV-FILTER based on dynamic rotation matrix to realize 3D world to 2D image projection.At the same time,Hungarian matching algorithm based on dichotomous matching is used in the process of obstacle matching(3)In order to show the effect of the algorithm proposed in this thesis intuitively,a set of fusion result visualization tool and semi-automatic annotation tool are developed.By synchronizing the radar raw data,visual raw data,radar clustering data and fusion results into the same time axis,and drawing the matching relationship of each frame,the current fusion effect can be easily observed.Based on the visualization tool and semi-automatic annotation tool,the tool can automatically generate the real value of obstacles at the current time by manually specifying the correct matching and matching at the current time.The results of the tool can be used to measure the accuracy of the fusion results,and the improvement effect of multiple fusion results can be compared.In this thesis,we use the simulation software carmaker to build a typical scene to verify the algorithm.At the same time,on the real vehicle platform,we collect the real automatic driving data of the actual road to verify the fusion algorithm.Simulation test and real vehicle test show that the fusion system proposed in this thesis has a great improvement in the accuracy,stability and continuity of the output results after fusion of camera sensor and millimeter wave radar sensor,and the fusion system has a high application value in the current automatic driving scene.