Submission information
Submission Number: 83
Submission ID: 117
Submission UUID: 31fe5d52-a3fa-4098-be7e-c6f8b19c866d
Submission URI: /form/project
Created: Wed, 01/06/2021 - 10:43
Completed: Wed, 01/06/2021 - 11:02
Changed: Sat, 08/21/2021 - 11:11
Remote IP address: 75.68.173.181
Submitted by: safwan wshah
Language: English
Is draft: No
Webform: Project
| Project Title | Utility poles Geo-Localization and Risk Estimation using Deep Learning |
|---|---|
| Program | Northeast |
| Project Leader | safwan wshah |
| safwan.wshah@uvm.edu | |
| Mobile Phone | 716-587-1594 |
| Work Phone | 716-587-1594 |
| Mentor(s) | safwan wshah |
| Student-facilitator(s) | Jamie Voynow |
| Mentee(s) | |
| Project Description | It is useful for communities and utility companies to maintain the quality of their utility poles. These poles are typically made from wood and can be damaged from various weather and environmental conditions. When a utility pole is damaged and falls, the entire community could face the consequences of power outages and dangerous electrical wires in/around the streets. Our goal is to leverage computing resources and deep learning to build geo-localization/assessment models for finding poles of poor quality/high likelihood of damage from street images. This tool will better prepare communities with poles at risk and the utility companies that support these systems to deal with the maintenance of these assets. The reason behind the superior performance of deep learning approaches is the availability of a large amount of labeled data. A more complete and sufficiently large dataset with geospatial information. In this project, we will use datasets provided by the Vermont Transportation department (VTrans). A previous collaboration between Vermont Artificial intelligent Lab and VTrans has introduced the Automotive Repository of Traffic Signs (ARTS), the largest dataset for U.S. traffic sign recognition. The dataset contains 65,152 sign annotations with spatial and geospatial information from 25,212 high-resolution images that include most of the MUTCD sign types and their GPS information. In this project, we will use the same annotation tools to label street pols. Labeling will require us to go through each image and specify poles attributes such as x, y coordinates by pixel and GPS location. Given a labeled utility pole, we can either develop a mechanism to calculate its angle of orientation, or we can also label this by hand. Once the dataset is in order, we can begin with the modeling phase. The most important underlying functionality of our model will be object detection specifically for utility poles. Deep learning is most well known for its contribution to object detection research, and we will be expanding upon this research domain with the addition of a utility pole assessment to our object detection system. Our final product will first detect utility poles and then assess quality/angle/positioning. If the performance of the detection model reaches satisfying performance we will enhance the detection model to include geolocalization. After the detection and geolocalization, we will estimate the falling risk. This can be done by either integrating this step in the detection model itself or implementing it as a post-processing step by calculating the pole angle from the detection model output. |
| Project Deliverables | - At least 1000 Vermont street labeled images/2000 object examples with geospatial utility poles and their attributes such as angle, type, coordinates, GPS position, etc. - Utility Poles Deep learning Object detection algorithms. - Algorithms for pole angle/quality assessment. - GIS maps containing detections of our utility pole models. |
| Project Deliverables | |
| Student Research Computing Facilitator Profile | James is a senior Undergrad student in CS department his research interest in machine learning to a wide range of computer vision and reinforcment learning applications. |
| Mentee Research Computing Profile | |
| Student Facilitator Programming Skill Level | Some hands-on experience |
| Mentee Programming Skill Level | |
| Project Institution | University of Vermont |
| Project Address | 590 Main street Burlington , Vermont. 05401 |
| Anchor Institution | NE-University of Vermont |
| Preferred Start Date | 01/11/2021 |
| Start as soon as possible. | No |
| Project Urgency | Already behind3Start date is flexible |
| Expected Project Duration (in months) | 3 |
| Launch Presentation | |
| Launch Presentation Date | 03/17/2021 |
| Wrap Presentation | |
| Wrap Presentation Date | 07/25/2021 |
| Project Milestones |
|
| Github Contributions | |
| Planned Portal Contributions (if any) | |
| Planned Publications (if any) | |
| What will the student learn? | - Basics Deep learning theory. - Deep learning tools and open-source software - Python package management - Large scale software development - Basics of GIS - GPU related programming |
| What will the mentee learn? | |
| What will the Cyberteam program learn from this project? | - Ideas related to pipelines in the computer version/deep learning process - The various challenges faced when using large scale open source software - Aspects of creating an object detection dataset from scratch |
| HPC resources needed to complete this project? | At least 500 of GPU hours and at least 50 GB of storage. |
| Notes | |
| What is the impact on the development of the principal discipline(s) of the project? | This research yields two major contributions to the object detection discipline: 1. The creation of a small object detection/GPS estimation dataset for utility poles 2. A baseline model of detection for this asset class |
| What is the impact on other disciplines? | Commercial and government operations concerned with accounting for their utility poles are able to use and expand upon this data in order to provide safety and maintenance. |
| Is there an impact physical resources that form infrastructure? | |
| Is there an impact on the development of human resources for research computing? | |
| Is there an impact on institutional resources that form infrastructure? | |
| Is there an impact on information resources that form infrastructure? | We have labeled over 1000 images containing almost 2000 objects of the utility pole asset class. |
| Is there an impact on technology transfer? | This data can be included in a larger dataset of all images containing labeled objects with bounding boxes and GPS location. |
| Is there an impact on society beyond science and technology? | This data will allow future technologies to be developed in order to protect communities from the dangers associated with fallen utility poles and their associated power outages, as well as the potential economic impact that these associated power outages can play in our world of post-pandemic online work infrastructure. |
| Lessons Learned | Python: - Using packages to interact with the operating/file system - Using packages to interact with image data - Using packages to interact with data structured hierarchically (.xml files) - Updating previously developed processes for GPU based systems Software Development: - Deploying code to GPU/Deep Green - Using version control - Working with open-source software - Building on top of massive software systems - Debugging package level errors - Package management with anaconda environments Bigger Picture: - Adhering to schedule is crucial for completing large projects - There are multiple ways to solve every problem, don't continue on a dead end just because you think you are too invested - Hand labeling data is a time consuming task - Asking for help is important - Working with open source software (especially software that is deprecated) is challenging (but also a great learning experience) |
| Overall results | There now exists a dataset for object detection and GPS estimation on the utility pole asset class. We have also manifested a model to estimate the number of poles in a given image. Future research endeavors can build upon this work, and hopefully be successful in implementing a working object detection system for this asset class. |