CrowdSight SDK: Guidelines, Camera Settings and Positioning

the further the person is from the camera, the smaller the resolution of the face in the image will be. Crowdsight performs optimally with faces of 150x150 pixels. The minimum face size for CrowdSight to perform in a meaningful way is 50x50 pixels (but the accuracy is lower).
In order to detect people from a distance, two things are required. Firstly, the capture resolution of the camera must be sufficiently large, 640x480 minimum is advised and 1080p is recommended. Secondly, optical zoom can be used in order to detect faces with sufficient resolution. As a general rule of thumb, when trying to record people from a distance larger then 5m you would require a camera with a zoom lens.
Camera Settings:
Depending on your camera manufacturer, it should be possible to configure camera parameters for best results. The most important this is to enable the anti-flicker for your region (50 or 60 Hrz) and to disable the automatic gain, white balance and focus control. This is a fundamental step, as the software will recognize any changes in the images as a change in the face, significantly affecting the classification and tracking results.
The ideal positioning of the camera would be aimed at a narrow space where people pass one by one, such as a doorway. The reason for this is that you can zoom in on say upper part of the door and capture each person's face one by one while they are entering or exiting. This allows for a higher face image resolution as opposed to the scenario where people can pass by alongside each other and thus a lesser level of zoom must be applied to process multiple people at once.
The best results are obtained when people are captured frontally. We advise to keep the angle between the face and the camera no more than 15 degrees. If you want to detect people at extreme angles, such as from cameras that are hanging on the ceiling you could attempt to attract their attention,for example by installing the camera next to an advertisement screen
The faces must be sufficiently and frontally lit. In case of nonlinear light patterns on the face or strong sideways illumination, CrowdSight will have more difficulty to correctly process the features of a person. Back illumination, such as a window behind a person, darkens the face feature and is thus less desirable. This can be counteracted to some extend by disabling automatic camera adjustment such as white-balance, focus and exposure controls, however this requires that the camera position or illumination does not change significantly. 
CrowdSight is able to work with infrared cameras that are using infrared LED emitters, this can be a solution in case a venue is poorly lit.
Have more questions? Submit a request


Please sign in to leave a comment.
Powered by Zendesk