Detection Types

The network to be used can be selected under Detection Types.

The generic network is always available and offers a couple of standard objects that it recognizes (for example, people, cars, etc.).

The number besides the dropdown represents a threshold in the range of 0.0 to 1.0. It represents a factor of confidence of how well an object has been detected. The higher the value the higher the probability the detection is correct. Object with a low confidence score can be ignored by using this threshold value. Depending on the quality of the images and the neural network this value can be adjusted to improve the overall tracking.

The list of available objects is shown under Detection Filter. This is the list of objects that are recognized by the neural network.

Detection Filter

Select which objects of you want to detect. In most cases, you would like to select one or two kind of objects in a video stream.

Tracking Types

The Tracking Types dropdown lets you choose what kind refined tracking to use inside of the detected bounding box. The button next to the drop down sends the new settings to the Object Tracker without having to use the global initialize button.

• AdaptiveCenter: Extracts color/luminance features around the center of the bounding box. A gravity force re-centers the tracking point toward the center of the bounding box over time to avoid drifting.

• AdaptiveTop: Extracts color/luminance features towards the top area of the bounding box. This is typically used when tracking persons as the arms and legs might move too much for a stable tracking, while the area of the chest and head are more stable and yield better tracking results. As for the tracker above, a gravity force re-centers the tracking point toward the center of the bounding box over time to avoid drifting.

• MagneticCenter: Works as AdaptiveCenter but when the object is not detected it uses the same algorithm of the Simple Tracking - Adaptive to try to follow the object; the downside is that the tracking is never lost, the operator has to manually turn off the tracking.

• MagneticTop: Works as AdaptiveTop but when the object is not detected it uses the same algorithm of the Simple Tracking - Adaptive to try to follow the object; the downside is that the tracking is never lost, the operator has to manually turn off the tracking.

• TrackerCenter: Extracts features around the center of the bounding box. A gravity force re-centers the tracking point toward the center of the bounding box over time to avoid drifting.

• TrackerTop: Extracts features towards the top area of the bounding box. A gravity force re-centers the tracking point toward the center of the bounding box over time to avoid drifting.

• Simple: Uses the center of the detected bounding box as tracked reference.

Detected object's bounding box

An example of a bounding box of a detected person object.

Tracking Types

This option tracks objects without the aid of a neural network: it uses the same algorithm that are used for the refined tracking in the "Detection and Tracking".

• Adaptive: Extracts color/luminance features around the clicked point.

• KCF: Extracts generic features around the clicked.

Filter Types

The filter types allow you to filter the tracked point over time. Several of them are available. The button next to the drop down sends the new settings to the Object Tracker without having to use the global initialize button.

• AdaptiveLowPass: Low pass filter that minimizes jitter and lag.

• MovingAverage: Calculates the average of the last points.

• ExponentialMovingAverage: Gives more weight to the most recent points. Similar to MovingAverage.
  
  

Size

Increasing the size of the filter helps to smooth noisy results.

Bounding Box Filter Size

The bounding boxes undergo a different filtering than the actual tracking point as it is more noisy and is subject to more extreme and sudden changes over time (for example, a person walking). You can configure a filter size for smoothing the bounding boxes for the tracking output. It is particularly useful when graphics need to be placed aligned to the bounding box. If you're using the Simple Tracking, the bounding box size slider is disable since there are no bounding boxes.