Search
  • Vineeth Veetil

Why Telemetric Sensors may not be enough - The Case for Computer Vision in Construction


Recently, several attempts have been made to address construction’s productivity challenge. There has been an impressive number of startups focused on construction tech, in Silicon Valley and beyond.

A large number of solutions are making it easy to digitize tasks once measured by a construction site operator or surveyor. Solutions such as PlanGrid and Fieldwire are seeing great adoption. These allow efficient collaboration between different stakeholders, such as supervisors, site managers and the crew by use of tablets and cellphones. These are great steps towards digitizing and managing the data, and making it easily accessible.

While above solutions are necessary, they still do not solve the root of the problem which is that the data collection itself is manual, expensive and very sparse. Other companies focus on capturing point-in-time snapshots of project progress using camera and LIDAR imagery. There are several cool drone-based solutions which can scan the construction site and create 3D point cloud maps. Often, analytics is built on top of these maps to understand construction progress. This type of visibility is a great improvement over manual solutions.

However, there are still challenges here. One, there is no visibility into the actual operations happening on the job site. These solutions make snapshots of the site using cameras, but there is no information on what happened in between, so no knobs to turn to improve productivity. Without actionable data regarding operations as well as resource utilization, there is little leverage for the stakeholders to improve operational efficiency. From McKinsey’s extensive reports and our own conversations with customers, the main challenges leading to inefficiencies are a lack of visibility into daily operations, which leads to poor resource management and utilization as well as late detection of issues. There is a need for detailed data regarding operations, such as the types of tasks performed and their duration. This can be correlated with project status to identify bottlenecks, reduce wastage and optimize resource utilization.

Another solution is IoT sensors that measure location using GNSS/GPS/SBAS, and other scalars such as temperature, pressure, humidity. However, IoT sensors do not have the cognitive capacity to measure complex manual tasks and unstructured tasks, or use contextual data effectively. This is a serious drawback. Also, there are material hurdles that IoT companies will have in real-world installations on job sites, including distributed devices, their maintenance, reliable communications networks, and the need for software integration from smaller sub contractors.

At Reflective AI, we are passionately building computer vision based software solutions to measure fine-grained operations in construction. Our product analyzes live video streams of operations, using deep learning-based action recognition. This data is analyzed to provide recommendations to cut operational costs based on insights, through better resource allocation, and early detection of issues leading to project delays or cost overruns.

A key advantage of such a system is the ability to leverage existing video infrastructure already installed on sites for security and other purposes. Since a large number of job sites have video cameras installed, this means lower set up cost compared to other solutions requiring investment in new hardware, complex setup, communications infrastructure, complex integration among multiple parties such as subcontractors and contractors, as well as expensive maintenance.

Technologically, this requires capability beyond simple image recognition and object detection. The solution requires understanding not only individual images but also the flow of information across frames in a sequence, called action recognition. Recently, breakthroughs have been made in this area, using advancements in deep learning specifically tailored for action recognition. This can then capture complex human and unstructured tasks and processes. We have hired some of the brightest minds in this space to solve this tough problem ! Stay tuned for updates.


0 views