Grantee Research Project Results
Final Report: CDD-SORT: A Next-Generation System to Detect Hazardous & RecyclableMaterials in Discarded C&D Debris
EPA Contract Number: 68HERD19C0027Title: CDD-SORT: A Next-Generation System to Detect Hazardous & RecyclableMaterials in Discarded C&D Debris
Investigators: Powell, Jon
Small Business: PTP Strategy, LLC.
EPA Contact: Richards, April
Phase: I
Project Period: May 1, 2019 through October 31, 2019
Project Amount: $99,995
RFA: Small Business Innovation Research (SBIR) - Phase I (2019) RFA Text | Recipients Lists
Research Category: Small Business Innovation Research (SBIR) , Small Business Innovation Research (SBIR): Phase 1 (2019) , SBIR – Sustainable Materials Management
Description:
The Construction and Demolition Debris Support Vector Machine Optical Recycling Technology (CDD-SORT) is our proposed system that detects hazardous, problematic, and recyclable components in discarded C&D debris at landfills and recycling facilities. During Phase I development, we named our software product Waste.AI. Our system and technology leverages a combination of emerging computer vision techniques, machine learning algorithms, and deep subject matter expertise to detect a wide range of materials commonly found in mixed C&D debris streams discarded to landfills./p>
Summary/Accomplishments (Outputs/Outcomes):
We created an initial target to characterize 10 different common C&D debris materials with our software. We developed a back-end algorithm that analyzes the photos by extracting and assessing various features. Key features included in the image classifier include color, shape, edge features, and user-specified information. We developed a front-end structure representing the graphical user interface (GUI) that enables the manipulation of files, manual tagging of images, and selection of training the model or analyzing a new photo. During development, we established a handful of design metrics to accomplish the goal of having an image classifier that can identify the proportional amount of various C&D debris materials in a photo of a pile of C&D debris – through this assessment, hazardous or recyclable components may be derived and used in the planned commercial application.
A working prototype was developed and hundreds of photos of individual loads of landfilled C&D debris were obtained and used to train the model. The prototype was trained on more than 200 images with a maximum-obtained accuracy of 76% across the 10 material categories. The accuracy was benchmarked against a human expert-labeled set of the same 200+ images. Accuracy varied widely and in general proportion with the complexity of the image, as expected. Reasonably high accuracy rates (>70%) were consistently achieved when three or fewer material categories were present, but error rates increased as image complexity increased. /p>
The labeling portion of the GUI segments the photographs into 144 segments in a 12x12 grid where each segment can be given a label by the software user. A key performance metric established during the project was processing speed of newly-loaded images, which we substantially addressed through the creation of a thumbnail loading system that enables rapid loading and only light reliance on computer memory. We established critical flow control between a splash screen, loading screen, project creation screen, and the main window.
Following initial front- and back-end development, the team built in dashboard analytics to output the classifier’s results into workable files (e.g., comma-separated value file indicating image information and classification results, bar charts, shaded images indicating classification results, and pie charts of multiple results). We built further improvements including a cropping feature to enable isolation of only C&D debris materials in a new image, which helps to exclude areas of interference including equipment, people, buildings, background/landscape, and areas of previously-placed waste materials.
Conclusions:
We successfully constructed a working prototype that includes all features necessary to classify an image of C&D debris across 10 common material categories. We trained our image classifier across hundreds of images of C&D debris, with results indicating an accuracy of up to 76% when compared to true (human-established) material amounts. Our system performed well with images containing a limited number of distinct material types, but displayed sometimes erratic and inaccurate results when classifying complex images containing >4 distinct material types. Customer discussions confirmed the economic benefit of our main value proposition (detecting prohibited materials), but most customers with whom we engaged expressed concern regarding the achievable accuracy of our system and the impact that in-field image capture equipment would have on operations.
Commercialization discussions were held with more than a dozen owners or operators of C&D debris landfills in the Southeast U.S. A few key patterns emerged during customer discussions. Encouragingly, nearly all potential customers concurred with our value proposition that a system that can satisfy spotting or screening needs at their site (spotting and screening refers to standard landfill practice whereby all inbound materials must be examined to assess whether any prohibited materials are delivered) would be valuable. The primary value as conveyed by the potential customers was that employees in the spotter position often take a long time to train and are sometimes subject to high turnover rates. Therefore a system that could allow those employees to be reallocated to other tasks (e.g., equipment operation), or possibly reducing hiring needs, could make the economics of our system attractive. Further, we found that our anticipated price point for deployment is consistent with most landfill operation budgets.
Three main customer objections emerged, which need to be addressed in the future to improve the odds for our technology to be commercialized. First, some customers indicated a desired accuracy of 90%+ to justify purchasing or renting an on-site image classification system. This accuracy outstrips that which the Waste.AI system can achieve at present. Second, most customers expressed concern about interference of image-capture equipment placed in the field with smooth operations. Namely, because cameras would need to be placed at or near the disposal working face (the location where trucks unload materials), several customers did not think there was a location that could be close enough to the waste to obtain good photos without interfering with truck traffic routes. Finally, some customers indicated they do not have recycling outlets for most types of C&D debris they receive, so these customers expressed skepticism about the benefit of our system to identify recyclable materials in inbound materials delivered to the landfill.
Journal Articles:
No journal articles submitted with this report: View all 1 publications for this projectThe perspectives, information and conclusions conveyed in research project abstracts, progress reports, final reports, journal abstracts and journal publications convey the viewpoints of the principal investigator and may not represent the views and policies of ORD and EPA. Conclusions drawn by the principal investigators have not been reviewed by the Agency.