Following describes how people counting works with SAFR to better understand how to optimize SAFR people counting results.  

For each person that enters the view, safr begins tracking the body.  If the face is within view safr will also associate that face with the respective body and perform recognition on the face to try to identify the person.  SAFR decides to do recognition based on quality thresholds.  If face is low quality (e.g. face too small, low contrast, blurry, etc), no recognition is attempted.  Once face image quality is sufficient, a biometric signature is created and db is searched for that identity.  If no match, person is identified as stranger, else identified as named person.

For each tracked person in the view of the camera, an event is started.  The event may be updated multiple times while open.  Once person leaves view of the camera the event is closed.  Person counting counts one person for each event. 

To ensure SAFR is counting people correctly, you should attempt to ensure that each person entering view of the camera generates one and only one event.