Bridges, Tunnels and other Transportation Structures
Full Session with Abstracts
The concept of citizen engineering has been previously introduced as a new paradigm in urban infrastructure monitoring that seeks to fill the gaps that exist within the assessment of urban infrastructure by proposing a cyber-human evaluation system. In this framework, non-expert citizens are engaged, trained and motivated to contribute data describing the urban infrastructure and its condition. A technical challenge facing this system is how to develop an efficient automated means of converting this crowd-sourced data into information, insights and decisions.
This work aims to leverage the recent advances in the field of computer vision and image recognition to develop a framework for image-based crowd-sourced monitoring of urban infrastructure and built environment. To this end, visual recognition models based on deep convolutional neural networks were built that can process urban infrastructure images and automatically detect and describe features of interest in them. Specifically, a dataset of the most common condition descriptors in urban environment was collected and annotated. These include infrastructure defects and degradation such as cracks, potholes, patches, and faded markings. To enhance the performance of the image recognition models, an auxiliary class of features including visual distractors from urban scene elements was added to the training set. Finally, the trained models were tested on user-generated samples and their performance was reported. Results demonstrate that the models are capable of classifying the images into the pertinent categories with excellent accuracy and therefore they can be integrated into the citizen engineering framework to automate the urban monitoring task.