Technologies



Camera


  • 4 MP, Full HD
  • 1/3” sensor
  • 60-170 degree fov camera lens
  • High sensitivity day/ night sensor
  • Dual bandpass IR coating
  • USB type C interface
  • Independent ISP

In-device Edge Hardware and Software


Edge processing driven by Intel hardware - core M3 based CPU, Integrated GPU and Intel Movidius (TM) VPU. Multiple deep learning/ neural network and vision computing based algorithms. Optimized for Intel platforms using Intel OpenVINO (TM) toolkit.

Recognition

  • Real time recognition - 70 ms
  • Min face size - 48 x 48 pixels
  • Recognition at 25-30 feet from camera
  • Out-of-plane rotation of -15 to +15 degrees
  • Configurable thresholds - face size, match value, number of matches - enables the software to be fine tuned to different deployments
  • 1k face vectors for easy shipment, storage and match in edge device while preserving privacy

Gender/ Age Classification

Determine gender based on frontal face image.

Determine age group based on frontal face image.

  • Infant 0-2 years
  • Child 3-12 years
  • Teenage 13-19 years
  • Young 20-35 years
  • Mid 36-55 years
  • Old 56+ years

Gaze tracking

Detect eyes from the frontal face image.

Determine if eyes are closed or fully open.

Continuous scale for intermediate eye positions.

Face Attributes

Determine following attributes from analysis of frontal face image

  • Use of eyewear
  • Use of headgear
  • Bald

Emotion

Determine if a person is smiling based on frontal face image.

Dominant Color of Clothing

Identify dominant color of clothing worn by a person whose face is detected. Color is based on the clothing for the upper part of the torso.

Face Tracking

Track same face across multiple frames even if the face is not detected in all frames to ensure all available faces are linked and made available for processing.

Track faces while crossing virtual boundaries to identify persons inside and outside a zone.

Count

Count people based on frontal face images include multiple features for accurate count.

  • Tracking when face is not detected
  • Recognition across frames to minimize duplication of count

Count based on human body profile - when frontal face image is not available.

Groups

Cluster persons in a group based on proximity in time (frames) and position within frames.

In-Device Interface


Send output of analytics algorithms to external programs on the device (eg media player).

The output is written in a message queue (RabbitMQ) which can be subscribed to by an external program and the events in the queue can be consumed to drive response such as change in content from the media player. Messages are delivered in real time (at a configurable frequency).

The output is also written in a CSV file which can be used for further processing.

Cloud based reporting


Use intuitive portal to access events, reports/ charts.

Analyze data aggregated across multiple locations; use different filters to zoom into specific parts of the operations.

Mobile


Receive alerts in real time on mobile to ensure they are acted upon rapidly including greeting a loyal customer or managing a slow moving queue.

Server Interface


The events generated by the analytics algorithm can also be sent to a cloud server through JSON/CSV which can be consumed by server software to generate charts/ reports for cross-device comparison and aggregation.

The JSON/ CSV can also be consumed by external server software for further processing.