Research On Error Characteristics And Influencing Factors Of Real-scene 3D Modeling Of Consumer Drone

With the development of consumer UAVs(Unmanned Aerial Vehicles),their advantages such as high flexibility,low cost,convenient data processing,strong environmental adaptability,and portability have won the favor of many scholars.The development of Sf M(Structure from motion)and MVS(Multi-view stereo)have led to the widespread application of consumer UAVs in fields such as 3D real scene modeling and terrain modeling in small areas.Consumer UAVs generally carry a single non metric camera,so the accuracy of 3D modeling can be affected by various factors.In addition,the 3D modeling errors based on UAV photogrammetry are often spatially variable and distributed,and RMSE(Root mean squared error)cannot reflect the spatial distribution characteristics of the errors.This paper uses the DJI-Phantom 4 Pro to obtain image data.Through designing several experiments such as flight scheme,control measurement,and using several evaluation indicators such as standard error,mean error,spatial distribution map of error,Moran’ I to analyze the overall characteristic changes of error.The paper hopes to provide scientific reference for 3D modeling of consumer UAVs which have same nature with DJI-Phantom 4Pro.The main research work of this paper is as follows:(1)In the condition of control free,the paper discusses the effect of camera angle and altitude.The single camera angle experiment proves that measuring accuracy can be improved without using a larger angle,while improving the spatial distribution of error requires a larger angle(>20°).The combined camera angle experiment shows that when the main camera angle is small,supplementing a few photos with other angles can improve accuracy,but the supplementing angle has no significant impact when the main camera angle is high.The single altitude experiment shows that within a certain altitude,the solution for improving accuracy is still based on the effect of oblique photogrammetry.The combined altitude experiment shows that it has a certain impact on the standard error,but has almost no impact on other accuracy evaluation indicators,and it is unnecessary to use the combined altitude when using oblique photogrammetry.(2)In the control measurement experiment,this chapter uses python scripts to repeat the bundle adjustment process,and explores error characteristic changes under different circumstances.The experiment shows that using a small number of GCPs(Ground control points)can rapidly improve the photogrammetry accuracy,and the accuracy tends to stabilize when the number of GCPs reaches a certain number.The same is true for the Moran’s I,but the number of GCPs required when the Moran’s I tends to be stable is more than the accuracy tends to be stable.The single camera experiment shows that the optimal GCP number required to stabilize when uses small camera angle is more than when uses high camera angle.Experiments on single altitude and combined camera angle show that the optimal GCP number is not sensitive to their changes.The combined altitude experiment shows that the necessity of combined altitude is not high.From the perspective of error spatial distribution,the accuracy of center distribution is the lowest,while the accuracy of edge distribution and spatial uniform distribution is better and their values are similar.However,the latter’s Moran’s I is lower and error is more uniform in spatial distribution.(3)During data processing,study the impact of camera model selection and modeling software selection on error features.By designing four camera model schemes,the experiment shows that when in condition of control free,the more complex the camera model is,the higher the accuracy improvement is.When has GCPs,different schemes do not significantly improve the accuracy.Considering the impact of complexity,the focal length,principal point,radial distortion,tangential distortion can be used as camera models to meet the needs.Select four mainstream modeling software to explore differences in operability.The results show that after adopting the optimization scheme,the accuracy has no significant difference.When using DJI-Phantom 4 Pro,DJI Terra has the shortest modeling time.Photoscan has the most adjustable parameters,suitable for exploring different influencing factors.