IntelliSite: Enabling safer and more efficient construction through video analytics and machine learning
Project summary and details for IntelliSite.
The COVID-19 contingency has highlighted the urgent need for site monitoring systems that enable measuring relative distances, effectively and accurately, among workers and plant equipment to ensure safe working conditions. Government guidelines have been useful, but they are not sufficient. For example, construction sites and manufacturing facilities have been shut down indefinitely due to large numbers of contagions among workers. The construction output fell by more than 40% in April (Office-for-National-Statistics). This underlines the need for accurate and inexpensive monitor systems that contribute to provide safe working conditions while maximizing throughput and productivity.
Areas of focus
Common camera-based site monitoring solutions focus on detecting damage due to crime and environmental hazards, such as intruders and fires. Other more capable systems enable limited object and change detection. Monitoring approaches that make use of Bluetooth devices as mobile phones and wearables have been proposed as well, but they reliability has not been proven yet (e.g. mobile tracking apps). Moreover, they require additional equipment, which increases costs and makes adoption more difficult.
This project proposes an innovative approach to develop the datasets necessary to enhance camera-based monitoring systems so as to improve significantly their current capabilities. This project can potentially deliver a qualitative step on the value that monitoring systems provide. For instance, by estimating activity efficiencies, safe distances, and identifying potential contagions. In addition, this project will provide a solution that will not require specialised equipment and will work similarly to current site monitoring systems.
16 months (January 2021–April 2022).
- Zest Consult Ltd
- Costain Ltd.
1. Generating synthetic datasets using 3D virtual scenes
- Genetic and procedural algorithms will be used to create a myriad of 3D scenes of construction sites using 3D models of workers, components, and plant equipment.
- Using virtual cameras, a synthetic dataset will be generated and automatically labelled considering different weather, lighting, and camera conditions.
2. Generating synthetic datasets using ML models
- Generative machine learning models such as GANs and autoencoders will be used to generate synthetic datasets from rea-life camera feeds on actual construction sites along with other methods such as transformations and distortions.
3. Object recognition, tracking, and photogrammetry
- State-of-the-art object recognition models will be used to recognise plant equipment and workers.
- Using feeds from various cameras, worker and equipment movements will be tracked and recorded.
- Photogrammetry will be used to measure relative distances, thus the distance among workers and potential risks can be detected, while equipment efficiency can be analysed.