Patch-Based Scene Segmentation for Construction Material Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current systems for monitoring the arrival and consumption of materials at construction sites are inefficient due to the complexity and non-standardized nature of video data collected by digital cameras, requiring specialized resources and being inadequate for processing large volumes of data.
Innovation Solution
A patch-based scene segmentation using neural networks is implemented, where data collection devices transmit data to processing computers that apply Convolutional Neural Networks (CNNs) to automatically identify and label materials in digital images, enabling efficient monitoring and tracking of materials, workers, and equipment.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If digital video cameras are deployed to collect videos and images from construction sites, then security monitoring and material tracking capabilities are improved, but the complexity of managing and processing the vast amounts of collected data increases
Solution Approach 1:
The patent segments the video data processing task by dividing frames into multiple patches or regions of interest. Each patch is independently analyzed by neural networks to identify specific materials (lumber, roofing, windows, doors), transforming a single complex processing task into multiple simpler parallel tasks that reduce overall system complexity.
Solution Approach 2:
The patent introduces neural networks as an intermediary between the video cameras and the management system. The neural networks automatically process and interpret the raw video data, extracting meaningful information about material arrival and consumption without requiring manual review, thus simplifying the data management workflow.
2Adaptability or versatility
If heterogeneous devices are used to collect data at construction sites, then device versatility and coverage are improved, but the difficulty of managing devices with different frame rates, time delays, and image resolutions increases
Solution Approach 1:
The patent employs neural networks that can process video data from multiple heterogeneous camera devices simultaneously. The neural network architecture is designed to handle variations in frame rates, resolutions, and timing by learning robust features that are invariant to these differences, allowing a single processing system to universally manage diverse device inputs.
3Measurement precision
If specialized and highly trained human resources are used to process collected video data, then processing accuracy is improved, but the cost and complexity of the system increases
Solution Approach 1:
The patent implements automated neural network-based processing that performs material identification and tracking without requiring specialized human reviewers. The system serves itself by automatically analyzing video frames, detecting materials, and generating reports, thereby eliminating the need for expensive specialized human resources while maintaining high accuracy.
4Measurement precision
If manual review of collected video data is performed, then monitoring accuracy is improved, but the time consumption and processing speed decrease
Solution Approach 1:
The patent replaces the mechanical process of manual video review with an automated neural network system. The neural networks process video frames computationally at high speed, automatically identifying materials and tracking their movement, thereby achieving both high accuracy and fast processing speeds that cannot be achieved through manual review.
Data Source
AI summary
A method and a system for patch-based scene segmentation using neural networks are presented. In an embodiment, a method comprises: using one or more computing devices, receiving a digital image comprising test image; using the one or more computing devices, creating, based on the test image, a plurality of grid patches; using the one or more computing devices, receiving a plurality of classifiers that have been trained to identify one or more materials of a plurality of materials; using the one or more computing devices, for each patch of the plurality of grid patches, labelling each pixel of a patch with a label obtained by applying, to the patch, one or more classifiers from the plurality of classifiers; using the one or more computing devices, generating, based on labels assigned to pixels of the plurality of grid patches, a grid of labels for the test image.