Business Objective
Our client is a leading oilfield services provider based in the US. The client had a manual process of monitoring safety events like checking if employees are wearing hard hats, monitoring unauthorized entry into restricted areas, and checking if pipes are aligned properly.
The client wanted a real-time video analytics platform to automate the site’s monitoring process, have an automated alert generation in place, and an AI-based entry log for restricted areas.
Challenges
- Camera placed far from the work floor and is susceptible to fog and light reflection
- Construction sites are cluttered with objects, and the occlusion of large objects is a major concern
- Need for a fast and efficient model that can process five frames per second on an HD video stream
Solution MethodologyÂ
- Analyzed client-provided video and other sourced datasets such as Visual Genome, CAVIAR
- Used YouTube videos and Google images to have sufficient training datasets
- For pipe alignment, used client-provided videos to train the model. For the hardhat and restricted area entry, used Visual Genome, CAVIAR, and other YouTube videos and Google images
- Extracted images using OpenCV and manually annotated them using the Labelling tool
- Trained and tested models for object detection and object segmentation
- Developed the capability to raise alerts in the form of light, alarm, email, or SMS through a flag in the system
- Deployed models with client surveillance feed coming in
- Pre-processed images using deep learning models to remove fog and glare leading to the detection of hard hats, multiple people tracking, and pipe alignment.
Business Impact
- Identified people, hardhats, and hardhat colors in different scenarios with 100% accuracy for person class and 98% for hardhat class
- Built a self-diagnostic module (SDM) that gives the working status of the video feed (checks for quality, light source detection)
