Skip to Main Content
 

Global Search Box

 
 
 
 

Files

File List


Liu_Xian_OhioLINK_3_12_19.pdf (9.39 MB)

ETD Abstract Container

Abstract Header

Feature Detection from Mobile LiDAR Using Deep Learning

Liu, Xian
http://orcid.org/0000-0001-5580-2121
http://rave.ohiolink.edu/etdc/view?acc_num=miami1552002747337465

Abstract Details

2019, Master of Computer Science, Miami University, Computer Science and Software Engineering.
Automated object detection from remotely sensed data in urban areas is a challenging task due to the complexity of urban scenes. Although recent advances in computer vision have shown that Convolutional Neural Networks (CNNs) are very well suited to this job, the design and training of such a CNN is still demanding and time consuming, due to the challenge of collecting a large and well annotated dataset and the specifi city of each task. In this study, we address problems in electric utility asset detection; including the problems of detecting electric utility boxes (for underground power cable) and also detecting the masts where above-ground power lines connect to houses. For the utility box problem we solve the box detection and oriented minimum-volume bounding box parameter prediction from mobile Light Detection and Ranging data (LiDAR), aided by Very High Resolution (VHR) orthophotos and geo-referenced videos. Our project can be divided into four stages: Since we are not aware of large-scale annotated LiDAR road scenery dataset that includes electric utility boxes, we annotated 196 positive electric utility box samples. Note that there are very well known LiDAR datasets collected for applications motivated by self-driving cars, but they do not include features relevant for the task of electric utility box detection. We trained and evaluated the effectiveness of using off-the-shelf pretrained models for classification in order to detect boxes by tuning them to predict the binary decision problem "does the current region of the image contain a box?". We also trained CNN to predict the rotation and location off set within each patch. Using such a classifier a sliding window approach, we are able to detect and catalog utility boxes in our test area.  We demonstrate the effectiveness of the Faster Regional Convolutional Neural Network (Faster R-CNN) framework to detect cars, houses and power-line attachments boxes from high quality georeferenced video imagery. The primary value of this contribution is to detect the masts and meter-boxes where power lines connect to houses although we also detect houses, poles, and cars because they provide potentially important contextual cues. We designed an automated system that can be con gured for easier model training and selection. The system is capable of comparing di erent data fusion, data augmentation, architecture and training options. We compared different models for data fusing, output parametrizations. Although road scene object detection using LiDAR is an intensely studied problem in autonomous cars, our trained models and automated system provides insight into the process of using a CNNs to recognize and catalog small and unconventional roadside objects. Our careful evaluation of a range of CNN models, data fusion approaches, and training schemes can be extended to the development of other high performance LiDAR based object detection systems.
John Femiani (Advisor)
Eric Bachmann (Committee Member)
Matthew Stephan (Committee Member)
64 p.
Computer Science
Deep learning, LiDAR, Feature Detection

Recommended Citations

Citations

  • Liu, X. (2019). Feature Detection from Mobile LiDAR Using Deep Learning [Master's thesis, Miami University]. OhioLINK Electronic Theses and Dissertations Center. http://rave.ohiolink.edu/etdc/view?acc_num=miami1552002747337465

    APA Style (7th edition)

  • Liu, Xian. Feature Detection from Mobile LiDAR Using Deep Learning. 2019. Miami University, Master's thesis. OhioLINK Electronic Theses and Dissertations Center, http://rave.ohiolink.edu/etdc/view?acc_num=miami1552002747337465.

    MLA Style (8th edition)

  • Liu, Xian. "Feature Detection from Mobile LiDAR Using Deep Learning." Master's thesis, Miami University, 2019. http://rave.ohiolink.edu/etdc/view?acc_num=miami1552002747337465

    Chicago Manual of Style (17th edition)

miami1552002747337465
709
© 2019, some rights reserved.Creative Commons License Feature Detection from Mobile LiDAR Using Deep Learning by Xian Liu is licensed under a Creative Commons Attribution-NonCommercial-ShareAlike 3.0 Unported License. Based on a work at etd.ohiolink.edu.
This open access ETD is published by Miami University and OhioLINK.