Computer Vision – Light Poles Inspection With AI Powered Vision System
Overview:
The government of Brazil conducts a survey that keeps the record of each and every electric pole installed across all different local regions. To prepare a database of all the poles installed, a manual inspection needs to be carried out. However, there have been inconsistencies in the results of the inspection and the process of the inspection itself is time-consuming. That is why Computer Vision – Light Poles Inspection With AI Powered Vision System was invented.
Client Requirements :
To identify multiple characteristics of three types (wooden and Metallic) of electric poles –
- Check the presence/absence of the street lights/illuminations
- Check the material of the poles (Concrete, Metallic, or Wooden)
- Check for the presence/absence of the distributor transformer fixed on the pole.
The inspection is completely manual. Operators are assigned to every region to conduct the inspection and record it manually. The average inspection time to completely inspect one pole is 25-30 seconds.
How AI Can Solve This Problem?An AI-powered vision system with a camera will be developed for the inspection of the poles.
The solution development journey is divided into 4 parts which are Image Acquisition, Machine Learning, Solution Deployment, and accuracy Improvement.
PORTABLE IMAGE ACQUISITION
An image acquisition software will be developed and fed into a tablet to acquire the images of the poles from different orientations and store them to QE®C (Qualitas EagleEye® Cloud). These tablets are given to a number of officials to take a set of images of the poles. The image acquisition part is the most crucial part of the journey as it helps to train the AI model in order to get correct accurate results.
MACHINE LEARNING
A solution is developed using the acquired images. Each type of pole (with and without the transformer and lights) is trained, with the help of a different set of images by making bounding boxes/circles around them i.e. also known as data annotation. This data annotation is done in QE®C (Qualitas EagleEye® Cloud) with a simple ‘point and click’ tool.
SOLUTION DEPLOYMENT
The trained model will then be installed on multiple devices (Tablets). These devices are portable vision inspection systems that are able to detect the material of the poles, lights, and transformer in real-time and display the results on the screen. Further, these results will be recorded into the database.
ACCURACY IMPROVEMENT
Deep Learning (DL) programs are created to train the machine vision system (Portable Tablets in this case) to understand the various untrained lights, transformers, and materials of the poles. The results will be reflected on the UI in real-time.
ConclusionPOC (Proof Of Concept) is conducted and the following conclusion is observed:
- False acceptance is reduced to 1 percent that would help our customers to reduce the recall rates.
- Inspection cycle time is reduced to less than one second that would help our client to increase delivery rates.
- Human intervention is reduced by 66 percent that translates to reduced labor and training costs