Automated gate control and facial recognition systems
- #Audio/video analytics
- #Computer vision
- #Deep learning
About the Client
A property management company from North America that manages elite-class real estate.
Business Challenge
Having an elite-class real estate comes with rigid security rules. Residents and guests want to feel comfortable and safe inside, while security should work perfectly and remain invisible. But often they need to provide an ID every time they enter the house, which is inconvenient and irritating. So, our customer wanted to implement a technology solution that would make the identification process seamless.
Solution Overview
The solution we developed offers a new user experience to the residents and the security service of elite real estate. Thanks to computer vision and deep learning, we armed security with a solution that detects and recognizes people, cars and various events (like a car pulling over). This way, residents and guests don’t need to show their IDs anymore to go through security checks to visit the facility.
The solution includes components for private parking management, guest parking management, backyard and entrance doors supervision. Moreover, it allows controlling more points with fewer people.
Project Description
Quantum has built a solution that consists of a data processing pipeline and components that implement specific features. As a basis, we used the Quantum video analytics platform that can be easily enriched with various features. After that, we have implemented functional components one by one: car detection, people recognition, license plate recognition, etc.
Security was one of the most important requirements, and that’s why it was important to run data processing locally. We recommended using NVIDIA Jetson TX2 for that, which is compact, energy-efficient, affordable and has an embedded GPU. Unfortunately, it’s computing power didn’t match the capacity offered by large data centers. In order, to meet the performance requirements, we adopted a data processing pipeline. For acceleration, we made strip frames and worked with images taken every five seconds instead of the 60 images per second that give modern cameras.
Let's discuss your idea!
Technological Details
Hardware:
- JetsonTX2 to deploy the model and use it in place
Software:
- Tensorflow for car detection and license plate recognition
- OpenCV for image preprocessing/postprocessing
- Dlib for license plate detection, face detection and facial landmarks recognition
- Darkflow Yolo for people detection
- SORT algorithm, built on basis of Kalman Filter, for tracking
- Python for the back end
