Optimized 3D Street Scene Reconstruction from Driving Recorder Images

Remote Sensing

Yongjun Zhang¹  Qian Li¹  Hongshu Lu³  Xinyi Liu¹  Xu Huang¹  Chao Song¹  Shan Huang¹  Jingyi Huang¹  Diego Gonzalez-Aguilera²  Gonzalo Pajares Martinsanz² 

[1] School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China; E-Mails:School of Remote Sensing and Information Engineering, Wuhan University, Wuhan 430079, China;;Electronic Science and Engineering, National University of Defence Technology, Changsha 410000, China; E-Mail:

关键词: street scene reconstruction;    driving recorder;    structure from motion;    outliers;    sparse 3D point clouds;    artificial intelligence;    classifier;   

DOI  :  10.3390/rs70709091

来源: mdpi

【 摘 要 】

The paper presents an automatic region detection based method to reconstruct street scenes from driving recorder images. The driving recorder in this paper is a dashboard camera that collects images while the motor vehicle is moving. An enormous number of moving vehicles are included in the collected data because the typical recorders are often mounted in the front of moving vehicles and face the forward direction, which can make matching points on vehicles and guardrails unreliable. Believing that utilizing these image data can reduce street scene reconstruction and updating costs because of their low price, wide use, and extensive shooting coverage, we therefore proposed a new method, which is called the Mask automatic detecting method, to improve the structure results from the motion reconstruction. Note that we define vehicle and guardrail regions as “mask” in this paper since the features on them should be masked out to avoid poor matches. After removing the feature points in our new method, the camera poses and sparse 3D points that are reconstructed with the remaining matches. Our contrast experiments with the typical pipeline of structure from motion (SfM) reconstruction methods, such as Photosynth and VisualSFM, demonstrated that the Mask decreased the root-mean-square error (RMSE) of the pairwise matching results, which led to more accurate recovering results from the camera-relative poses. Removing features from the Mask also increased the accuracy of point clouds by nearly 30%–40% and corrected the problems of the typical methods on repeatedly reconstructing several buildings when there was only one target building.

【 授权许可】

CC BY   
© 2015 by the authors; licensee MDPI, Basel, Switzerland.