Unsupervised Feature Learning for Objects of Interest Detection in Cluttered Construction Roof Site Images

Abstract

In the roof contracting industry, safety violations continuously lead to fall injuries and fatalities. Occupational Safety and Health Administration (OSHA) suggests standard protective measures, but they are often not followed due to factors such as tight budget and lack of training. To alleviate this situation, we propose to develop a system that can automatically check the compliance of fall protection standards through machine vision and learning techniques to exploit day-to-day site images collected by the surveillance videos and site engineers. As an initial effort, this paper focuses on evaluation of an unsupervised feature learning and image classification method i.e., Convolutional Neural Networks (CNN) to detect objects of interest (roofs, roofers, guardrails, and personal fall arrest systems) in a large number of unordered and cluttered construction site images. To isolate different objects, we initially segment each image using Gaussian Mixture Model (GMM) and pass the resulting segments as input into CNN. This enhances the feature distinction between different objects and augments the inter-class variability. Then, we extract large feature sets in a hierarchical manner and classify images based on the acquired object features. Experiments results signify the promising performance of the CNN method in terms of accuracy. This research demonstrates potential of this method and paves the way towards applying it in the next research development required to achieve our ultimate goal.