What to expect from this technology

At present Accuware Sentinel can: ​

At any instant a Person ID is assigned to each individual ​​on the field of view of each video camera. ​The assigned IDs generally “stick” to the same individual for the time the individual remains in the field of view of the video camera and if the scene is not overcrowded. Despite this, it must be kept in mind that at present:

  • at different instants, the same Person ID could be assigned to different individuals AND at different instants, different Person IDs could be assigned to the same individual. The Person ID  “stick” to the same individual as long as he/she remains in the field of view of the video camera. If individuals leave the scene and later re-enter it, new Person IDs may be assigned to them if their appearance is somewhat different because of varying lighting conditions.
  • at the same instant, inside 2 or more video streams, there could be different individuals identified by the same Person ID. At present Person IDs are not shared between different video streams.​
  • some individuals are not detected. This happens for example when they are far away or they are not completely inside the field of view of the video camera. ​Recognizing people in a busy scene is challenging. People on screen must be at least 30 PX wide to be recognized. Occlusion (people obscured behind other people) and groups where individuals cannot be picked apart are real possibilities. People counting applications must take into account that these counts are approximate.
  • there could be false positives. Some anomalies should be expected, such as recognizing a human shape displayed on a poster as if it were an actual individual in a scene.
The goodness of the results provided by the Accuware Sentinel algorithm are very dependent on the video streams. Below we grouped the most frequently asked questions related to the accuracy.

How well does Accuware Sentinel handle low/poor lighting?

A poor lighting of the scene affects the individuals detection if the contours/silhouettes are not apparent enough. So, in general the individuals are still detected but a poor lighting could directly affect the ID assignment if the colors tend to be too grayish.

How well does the system handle large crowds?

We have decent performances in crowded scenes (many individuals on the scene). Despite this, if the crowd is really dense, and individuals are too close to each other, the system won’t work well and there will be a lot of missed detections and IDs switches. Individuals are detected by the algorithm when a “reasonable” part of the body is visible. In a very crowded scene, where only heads are not overlapping, the detection can be difficult.

Does the system handle multiple cameras looking at the same point from different angles?

No, not at present (but we are working on this).

Can the system recognize that it is looking at the same individual from different cameras at the same time? ​

No, not at present (but we are working on this).

If an individual returns on a different day, can the camera detect that it is a return individual?

No, because at present the algorithm does not perform facial recognition (and in general the quality and the location of the average cameras do not make possible facial recognition). The assigned IDs generally “stick” to the same individual as long as he/she remains in the field of view of the video camera and assuming the scene is not overcrowded.

How well are individuals detected when not facing the camera, when in a large crowd?

Individuals are detected even if they are not facing the camera, but the large crowd can be problematic as explained previously.

How far away from the camera do individuals stop being detected?

It depends on the video resolution and on the location of the camera. So, in the end, it depends on the specific video stream.

How well are individuals that aren’t moving detected?

We are filtering non moving individuals and objects. Anyway the algorithm is able to track individuals that: walk, stop and then start walking again.

The system currently has about a 10 minute delay of data, how much does this increase by with large crowds?

​​It does not increase because we can manage the computing capacity in real-time.

Are individuals that walk behind objects detected? ​​​​​​

Yes, but it depends on how big is the obstacle and for how long an individual disappears from the field of view.

Will individuals that aren’t completely in the camera’s field of view be detected eventually? ​​​​​​

The algorithm needs a “reasonable” part of the body, but for instance, the algorithm is able to detect individual even if only the upper body is being seen.

How many individuals can be detected at one time from a single camera?

​There is not a limit. It all depends on how crowded is the scene.

How far individuals can be in order to be detected?

The maximum detection radius depends on how the camera is mounted and the resolution of the camera itself. People need to be at least 30 pixels tall in the video stream in order to be detected.