Unmanned Computer Vision algorithm
Infrastructure inspections are now performed by expert technicians due to their critical nature. Currently, unmanned drone inspections are at the forefront of the industry, employing highly specialised hardware to record images and videos of the studied structures that are later inspected by these technicians.
On one hand, these advances have eased the inspection process while saving valuable time; especially during the inspection of difficult-to-access structures where a higher risk is taken by construction workers. On the other hand, technicians spend lots of time reviewing video feeds since no preprocessing of the data is done beforehand; this may be many hours of videos per day, depending on the demand of the company’s clients. This is also a highly monotonous task, which means that even expert technicians may miss some defects due to the difficulty of maintaining attention in this task.
Creating an intelligent system that filters the videos and searches for defects prior to the expert’s inspection would save time, since the technicians would need to review less footage, saving time and costs and scaling their business to more clients, while also improving accuracy.
Bedrock designed and developed an end-to-end Computer Vision solution for the automatic detection of defects, employing the latest advancements in transformer-based neural networks.
This solution would provide the clients with a solution to pre-label their video feeds: the AI model would detect the presence of different kinds of defects and label them throughout the video timeline.
A holistic solution was developed on the Microsoft Azure cloud platform, following MLOps best practices to ensure robustness and flexibility. Furthermore, separate client storages ensure privacy concerns regarding where their data is kept.
For the model training, Bedrock partnered with a local company for using unmanned aerial drones to record videos at real locations. This guaranteed model performance to be closer to real-word applications our clients would require it for.
The Computer Vision cloud solution provided a highly-accurate AI model that could detect defects under the filming conditions of the inspection process.
An additional logic was added to the final model in the case of underperforming detection results; in that case, new data inputs would differ from previous recorded videos, and a model re-training would adapt it accordingly.