Solution highlights

  • Object detection and face recognition
  • Accurate face mask identification
  • Body temperature detection
  • Real-time social distancing monitoring
  • Advanced people counter module
  • Instant alerting with violation details

Challenge: Detect, identify, measure

The client needed to create an enterprise-grade solution able to count the number of employees in every particular area as well as identify whether people wear masks correctly and observe social distancing rules.

To meet this business need, we used our computer vision and deep learning competence. The SCRUM methodology was chosen to easily adapt to constantly changing requirements and innovative project nature, while ensuring full process transparency.

“Project complexity required a highly competent, cross-functional team. We assembled frontend and backend engineers, DL experts, Python developers, and QA specialists with 50+ years of cumulative team experience.”

— Anatoliy, Project Manager

Delivering a robust employee safety system

Oxagile’s decade-long experience in developing computer vision solutions proved instrumental in building an innovative system able to control whether COVID-19 restrictions are met within the enterprise.

Precise face detection

  • Detection of multiple faces despite occlusions
  • Accurate employee identification (name and surname)
  • Face image processing speed: 15 ms
  • Unified data processing component that allows adapting to new case scenarios
  • Optimal “accuracy vs computing requirements” ratio for minimized TCO

Face mask recognition

Our engineers worked with the MAFA dataset (30,811 images and 35,806 masked faces) to immediately detect mask wearing violations, while dealing with the most complex real-life scenarios. The delivered solution detects:

  • Correct and incorrect mask wearing
  • Uncovered faces (no masks)
  • Occluded face instead of a mask
  • Masks that don’t cover the mouth or nose
  • Full face mask wearing

Body temperature determination

  • Thermal cameras implementation
  • Smooth processing of a 16-bit thermal images
  • Using non-standard dataset with grayscale thermal images
  • Simultaneous utilization of RGB and a thermal camera
  • Temperature determination accuracy: +/- 0.3 °C (~0.5 °F)

Object detection and tracking

  • Real-time detection of multiple objects
  • Advanced counter of people entering and exiting the building
  • Social distancing monitoring (minimum safe distance, room capacity)
  • Immediate rule violation alerts highlighting every particular violator in person
  • Video post-processing to increase detection and recognition accuracy

Expert help with your solution

Leverage our computer vision knowledge to build a custom solution from scratch or enhance an existing one with cutting-edge functionality like face identification, object tracking, behavior analysis, and more. Contact us describing your requirements, and we’ll help you find the balance between cost, performance, and business value.

Pushing scalability with edge devices

To scale the solution’s capacity on Nvidia Jetson edge devices, we leveraged the C++ programming language. This allowed us to easily handle the streams from multiple cameras and smoothly perform all detections and calculations.

Depending on modules used, every edge device was connected to up to seven cameras. And the number of edge devices used in one tenant (web application) was unlimited.

Capitalizing on deep learning expertise

Oxagile’s growing DL expertise across the production department helped us dive deeper into project requirements and implement them strongly at the business logic level.

We used the best practices around computer vision to ensure unmatched accuracy of video and image analysis — despite highly compressed surveillance videos, moving objects, various face angles, and shaky camera footage.

Our proprietary computer vision know-hows were instrumental in accelerating data analysis as well as ensuring the optimal balance between image processing accuracy and system performance.

Face mask recognition accuracy

Face image processing speed

Temperature determination accuracy