AI Confidence Settings – Rhombus Support

Topics included in this article:

Overview
AI Confidence Thresholds
- High Confidence Setting
- Low Confidence Setting
Adjusting AI Confidence
- Increasing Confidence
- Decreasing Confidence
Environmental Impacts to AI Performance
Troubleshooting AI
Helpful Links
Contact Support or Sales

Overview

AI confidence thresholds determine how certain a Rhombus camera must be before identifying a person, face, vehicle, or license plate. Adjusting these thresholds can help balance missed detections and false positives based on the camera's environment and positioning. This article explains how AI confidence thresholds work, how to adjust them, and how to troubleshoot common AI detection issues.

AI Confidence Thresholds

In AI detection models, confidence refers to the AI's certainty that an identification is correct. Confidence thresholds determine how certain the AI must be before identifying an object as a human or vehicle. Adjusting these thresholds can help balance false positives and false negatives.

High Confidence: The model uses a stricter threshold when identifying humans or vehicles, reducing incorrect identifications but increasing the risk of missing valid targets.
Low Confidence: The model uses a less strict threshold when identifying humans or vehicles, reducing the risk of missing valid targets but increasing the risk of incorrect identifications.

High Confidence Setting

When the confidence threshold is set high, the model requires a higher level of certainty before identifying a human or vehicle. Setting the threshold too high can result in missed detections when the model does not have enough information to confidently make an identification.

Example: Imagine a security camera is configured to detect humans with a confidence threshold of 95%. The system will only identify a person if it is at least 95% confident that the detected movement matches its model for human movement. If the person is partially obscured or the image quality is poor, the model may not reach the required confidence threshold, resulting in a missed detection.

Low Confidence Setting

When the confidence threshold is set low, the model requires less certainty before identifying a human or vehicle. Setting the threshold too low can result in false positives because the model may make an identification without enough information to accurately classify the detected object.

Example: Imagine a security camera is configured to detect humans with a confidence threshold of 50%. The system will identify a person if it is at least 50% confident that the detected movement matches its model for human movement. Because the required confidence is lower, objects such as moving tree branches or birds may be incorrectly identified as people, resulting in more false detections and alerts.

Adjusting AI Confidence

Finding the appropriate confidence threshold for a device may require some adjustment. While changing the confidence threshold can improve AI detection performance, external factors such as ambient lighting, camera positioning, and camera angle can also significantly impact detection accuracy.

AI confidence thresholds can be adjusted for the following event types:

Human Detections
Face Detections
Vehicle Detections
License Plate Detections

Note: For the AI model to identify a face, it must first identify a human in the frame. Similarly, the model must first identify a vehicle before it can identify a license plate.

Increasing Confidence

Increasing the confidence threshold requires the AI model to have greater certainty before making an identification, which can reduce false positives. However, setting the threshold too high can cause the model to filter out valid event detections.

Decreasing Confidence

If one decreases the confidence threshold, the AI will be more lax with its identification criteria to reduce the number of missed events. This increases the number of detections being made, however, going too far in this direction can lead to false positives.

Environmental Impacts to AI Performance

Successful AI detections rely on an appropriate environmental setup. All camera setups are susceptible to environmental factors that can reduce the effectiveness of the AI model. Each detection type may require a different environmental configuration.

One key environmental factor is external lighting. Adequate lighting is vital for cameras to identify objects such as people, faces, vehicles, or license plates. Without sufficient lighting, our cameras may not be able to make these identifications. Additionally, the height and angle at which a camera is mounted can also affect detection accuracy. Click here for more details.

Troubleshooting AI

Note: AI detections are heavily dependent on camera positioning. Before troubleshooting AI, ensure your camera is mounted and angled appropriately for the event type you try to detect. For more details on optimal camera positioning for specific event types, click here.

Issue	Solution
There are false positives. One event type is picking up detections when that event is not occurring.	The confidence threshold for that event type may be set too low, causing the model to identify events incorrectly. Increase the confidence threshold for that event type by 0.2 or 0.3 increments and monitor improvement.
No Human Movement is being detected.	If all environmental factors are set up for optimal performance, decrease the Human Movement confidence threshold by 0.2 or 0.3 increments until the system begins to capture human movement. Best Practices for AI Optics and Object Distances for Analytics
Facial recognition is not picking up faces.	If all environmental factors are set up for optimal facial recognition, ensure Human Movement is properly being captured. If there is associated Human Movement, decrease the Facial Recognition confidence threshold by 0.2 or 0.3 increments and monitor for an increase in face detections by the system. Managing Facial Recognition
No Vehicle Movement is being detected.	If all environmental factors are set up for optimal performance, decrease the Vehicle Movement confidence threshold by 0.2 or 0.3 increments until the system begins to capture vehicle movement. Best Practices for AI Optics and Object Distances for Analytics
No license plates are being detected. (LPR)	If all environmental factors are set up for optimal license plate recognition, ensure that Vehicle Movement is being detected. If there is associated Vehicle Movement, decrease the License Plate confidence threshold by 0.2 or 0.3 increments and monitor for an increase in license plate detections by the system. Managing License Plate Recognition (LPR)

Helpful Links

Contact Support or Sales

Have more questions? Contact Rhombus Support at +1 (877) 746-6797 option 2 or support@rhombus.com.

Interested in learning more? Contact Rhombus Sales at +1 (877) 746-6797 option 1 or sales@rhombus.com.