Detection Settings
Detection settings provide the user with tools to configure preferences regarding the AI detection mechanism.
Navigating to the Settings
After selecting a task under the Task Dashboard, you will be directed to that individual task's interface. Click Settings in the top right corner to enter the System Settings, then click Detection.
AI Confidence Threshold
The confidence threshold is the minimum certainty level required for the AI to report an object. Setting a high threshold acts like a strict filter: it ensures accuracy (fewer false alarms) but risks being "too picky" and overlooking actual objects. Conversely, a low threshold captures more objects but increases the risk of mistaking noise for a target.
Each ontology can have a different AI confidence threshold. Click Edit on the lower right corner and configure each individual ontology's threshold.
Object Size Filter
This filter defines the acceptable dimensions (width% and height%) for a detection to be considered valid. Any bounding box falling outside these minimum or maximum boundaries is discarded as noise or irrelevant data.
This feature helps in perspective control and helps account for depth; if an object is "too big," it might be too close to the camera to be identified correctly, or it might be an insect crawling across the lens. And if you are monitoring a highway for cars, you can filter out anything the size of a squirrel to prevent "phantom" detections.
Each ontology can have a different object size filter. Click Edit on the lower right corner and configure each individual ontology's height and width range. As a reference, a ruler view has been added across the stream. You can use this ruler to estimate the size of the bounding boxes. In this case, if we just want to monitor the intersection, a car height and width range of around 3-30% is sufficient.
Once satisfied, click Save to save your details.
Image Quality Filter
This feature functions to filter camera input based on image quality to reduce false alarms. This setting is turned off by default. Enabling the image quality filter lets the model ignore image frames that do not meet the criteria set by the user below:
Minimum Input Resolution: Low resolution reduces visual detail and harms detection. Here, the user can set the minimum input resolution that the model will accept.
FFT Blur: This technique uses Fast Fourier Transform (FFT) to calculate how much sharp edges there are in the image. Therefore a low FFT blur means less edges and more blur.
BRISQUE: BRISQUE is based on the principle that undistorted images follow certain mathematical properties in their pixel intensities, where natural images tend to have a Gaussian-like distribution in their luminance. When noise, compression artifacts, or motion blur are introduced, these statistics shift away from the "natural" model. The scoring for BRISQUE is the inverse of FFT, measuring the amount of distortion.
Caching Inference
When consecutive frames are highly similar, the system skips AI inference for the subsequent frame to reduce computational load. This is highly effective for stationary cameras in environments with minimal scene changes.
When turned on, the user can toggle the following functionalities:
Sub-Sampling: Determines the density of pixels analyzed when comparing frames. A higher value (e.g., 16 px) skips more pixels to increase speed but may miss very small movements. A lower value (e.g., 1 px) checks every pixel for maximum precision. The default value of 8 px indicates that the system only looks at the pixel every 8 steps. It checks the pixel at (0,0) then skips to (0,8), (0, 16) and so on to check for similarity.
Similarity Threshold: The percentage of change required between frames to trigger a new AI inference. Lower % means high sensitivity; even slight movements (like leaves rustling) will trigger the AI. Higher % means low sensitivity; the system ignores minor "noise" and only runs the AI when significant movement occurs.
Anonymization
The system can automatically blur faces and license plates to protect personal privacy. When anonymization is enabled, the system blurs faces and license plates in the event snapshot to protect personal privacy and comply with regulations.
AI Model Version
This section allows you to setup the AI detection model version.
Last updated