HEA BLO INS Machine learning Automation-at-Scale MAS

How machine learning automates telecom asset inspection

Providing quality products for Pix4D users means learning about and using the best of modern technology.

Pix4D strives to be a company that innovates and develops next-generation technology and techniques for its users. As a market leader in photogrammetry, the company has been branching out into new, specific industries in response to increasing demand and seeing drone-related solutions Pix4D could provide. One example of this is our machine learning software.



Pix4Dinspect is a prime example of Pix4D using machine learning, partnered with the Pix4Dscan app. Pix4Dscan has specific flight plans for cell tower inspections, whilst Pix4Dinspect uses artificial intelligence (AI) and machine learning (ML) to speed up the analysis of point clouds generated from data gathered by Pix4Dscan flights. ML simplifies the inspection of 3D digital twins of telecommunication towers, removing the need to manually detect and measure equipment components. The resulting digital tower isn’t just a replica of reality, but an intelligent digital twin thanks to the use of machine learning.

“Pix4D offers unique opportunities in this space of AI and machine learning that makes us very innovative and competitive” - Andrea Dotti, Core Analytics Team Manager at Pix4D

The virtuous cycle of AI

Andrea Dotti explains that the process for using AI and machine learning at Pix4D follows the best practices in the industry. In the words of Andrew Ng, Professor of Computer Science at Stanford University, California: “We’re making this analogy that AI is the new electricity. Electricity transformed industries: agriculture, transportation, communication, manufacturing”. The potential for ML and AI is industry-changing. One of the basic but most powerful concepts in the machine learning world is the “virtuous cycle of AI”, where the process of AI developing is part of a repetitive cycle moving between the product, users and data collection.

VIS Product data user MAS (1)
The virtuous cycle of AI

Andrea explains that humans must validate machine learning results, to tailor the purpose of the technology in order to constantly improve it so that the machine learning will progress. This is how the virtuous cycle of Artificial Intelligence works. The result of this workflow is 100% focus on users and their experiences, which constantly improves the quality of a product.

However, outperforming at each cycle of the ML results is not an easy process. In order to increase the accuracy of the software and its results, more data needs to be gathered. As more data is injected into the pipeline, the machine learning models are trained to see more variations of what is of interest, hence improving their prediction and analysis capabilities - for example with Pix4Dinspect, learning what a telecommunication antenna looks like and the position it is fixed in. As reflected by the virtuous cycle of AI, the more this process of increasing the available data is repeated, the more the accuracy of machine learning improves. This is immediately clear in Pix4Dinspect, as thanks to this iterative process the downtilt angle (the most important parameter) can now be measured with extreme precision.

Automation at scale: the challenge of modern-AI business

IMA BLO INS Machine learning Automation-at-Scale MAS (1)
Andrea describes introducing this technology workflow in a modern business: asset preparation, machine learning analysis, and finally, roll out to users

Pix4Dinspect capitalizes on ML benefits: the cloud platform has been engineered to use machine learning to automatically recognize antennas using the images and data captured by Pix4Dscan. Pix4D is capable of implementing this ML solution in its software without requiring third-party algorithms, instead completing the integration with in-house software engineers. Throughout the process 2D and 3D models are seamlessly connected to AI, from the drone capturing data to analytics. As a result an immense workload is removed from the client’s inspection process. The deep learning algorithm for antenna recognition is able to detect antenna patterns in the point cloud and compute their position, orientation, and dimensions. Pix4Dinspect automatically detects antennas and reports on these specific features:

  • Antenna type: panel or microwave.
  • Downtilt: Downward angle with respect to the gravity field lines. Values range from -90 to +90 degrees.
  • Azimuth: Yaw angle with respect to the true North. Values range from 0 to 360 degrees.
  • Plumb: Roll angle (side tilt) with respect to the gravity field lines. Values range from -180 to +180 degrees.
  • Width: Width of the antenna.
  • Height: Height of the antenna.
  • Depth: Depth of the antenna.
  • Altitude (CL): Altitude from the center line of the antenna to the ground

Automatically producing a report, machine learning algorithms dramatically reduce the burden of work on users because users no longer need to manually label antennas and can instead concentrate on only adding project-specific annotations which often requires just a single click. The information extracted by ML is not only useful for maintenance, inspection and asset verification, but also the global roll out of 5G, which requires up-to-date information about the status of cell towers. The identification of these features does not happen by magic, but by the ML system following three key stages, which Andrea explains like this:


Key stages of machine learning at Pix4D

  • Use state-of-the-art ML to segment antennas in images from Pix4Dscan. This eventually is limited by the capacity of 2D visualization.

  • The 2D segmentation masks are projected into 3D space, and the resulting 3D points are clustered and the antenna detections refined.

  • 3D spatial symmetries are used to find high precision antenna poses, giving information like width, depth and azimuth.

The impact of machine learning on assets inspection

As a result of the incorporation of ML into the workflow of Pix4Dinspect, Pix4D has dramatically reduced the inspection time of telecommunication towers.

A manual antenna detection can take half a day with 10 degrees of accuracy. The ML solution provided by Pix4D is automatic, happening during normal processing of the tower for a 3D model, has a downtilt precision of up to 1 degree. Not only is the software faster than manual inspection but it has greater accuracy. Included with that is the fact that Pix4Dscan needs only 20 minutes flight time to complete a standard cell tower inspection, whilst Pix4Dinspect takes 2 hours to process it.

This does not change regardless of how many assets are being processed at the same time due to the use of cloud resources, so customers do not challenge their own IT infrastructure.

As many towers as necessary can simultaneously be processed, making the total inspection time a matter of a few hours rather than several days, which was proven when SkyVue Solutions used Pix4Dscan and Pix4Dinspect to inspect a cell tower in Douala, Cameroon.

SCR INS Antenna detection MAS
The Pix4Dinspect interface

This ingenious pipeline is leveraging unique capabilities that are only available at Pix4D, since the entire pipeline is curated in-house, from data acquisition to the processing of images and the representation of results in reports. Pix4D implements a tight and unique data processing workflow that can use the best in photogrammetry, computer vision and deep learning algorithms to provide users with a complete, precise and cost effective solution.

Scale your visual inspections with AI

Related articles