Welcome to the first in our three-part series on installing and optimizing cameras for DISPL’s AI-powered audience analytics. In this article, we’ll give you a high-level look at how our face detection algorithm works, the overall workflow from camera capture to analytics, and why factors like lighting, angle, and camera quality are so critical.
1. What Is Audience Analytics and Why Use It?
DISPL’s Audience Analytics (also known as face detection or face analytics) is designed to analyze when people are in front of a screen or kiosk, measure how long they look at it, and capture demographic attributes (e.g., approximate age, gender). This helps you:
- Optimize content based on real-time audience data.
- Understand store traffic by measuring how many unique visitors engage with the screen.
- Gain insights into dwell time and how effectively your on-screen content grabs attention.
Importantly, the system does this in a privacy-conscious way:
- It captures only face descriptors (mathematical patterns) without storing personal images.
- It does not perform liveness checks, so printed or digital face images can also be detected if they appear in the camera’s view.
2. Core AI Algorithm: Detection, Tracking, and Analysis
Our solution’s AI pipeline works in three main steps:
-
Face Detection
- The system locates any face-like region in the camera feed.
- Each found face is assigned a bounding box, and the algorithm checks whether the face meets the minimum pixel size requirement (typically at least 60 pixels in height).
-
Face Tracking
- Once detected, the same face is tracked over time (in consecutive frames).
- A Track ID is generated to follow it as it moves. This ensures the system knows when a visitor is still in view or has left.
-
Face Analysis (or Person ID Assignment)
- If enabled, the system attempts to generate a face descriptor to recognize that individual again if they reappear.
- The descriptor is a unique numerical representation of a face—no raw image data is permanently stored.
- High Sensitivity Mode can be turned on to allow recognition from partial or profile views; otherwise, the algorithm focuses on more frontal images to reduce false matches.
Key Note: For best results, a visitor’s face should be visible for at least 2–5 seconds. Quick glances (under 2 seconds) might not be detected or included in analytics.
3. Typical Data Flow: From Camera to Analytics
Below is an example of how the process works in a store or retail setting:
-
Camera Capture
- A USB camera (or built-in webcam) feeds live video at 720p or higher, ideally at 25–30 FPS.
- Our software continuously analyzes each frame, looking for faces.
-
Local AI Processing
- On the local device (player/PC/kiosk), the AI engine detects and tracks faces in real time.
- If the system is configured for offline usage, these insights remain local until a network connection is available.
-
Data Aggregation
- The detected face(s) generate events such as “New visitor,” “Returning visitor,” “Face left camera view.”
- The analytics engine records dwell time, approximate demographics, and unique Person IDs (where enabled).
-
Reporting & Visualization
- Summaries of these events (e.g., how many unique visitors, average dwell time, or demographics) can be sent to the DISPL Content Management System (CMS) and stored in the cloud.
- The data is then available in dashboards or analytics reports to help refine your retail media strategy.
Tip: If multiple sensors (cameras) operate on the same network, they can share data so the same visitor isn’t double-counted when they move between nearby screens—though there’s a brief window before synchronization.
4. Why Lighting, Angles, and Camera Quality Matter
4.1 Lighting
- Well-lit environments ensure the camera can capture enough detail to detect and classify faces.
- Sudden changes in brightness or strong backlighting can cause faces to appear too dark or washed out.
- In dim conditions, consider adding additional ambient light or IR illumination (though IR may slightly decrease accuracy).
4.2 Angle
- The system performs best when the face is within ±30° of a straight-on view.
- Extreme angles (looking up from the floor or down from the ceiling) distort facial features and reduce detection accuracy.
- If you must mount the camera higher (e.g., above 2.5 meters), use a long-focus lens or position it farther away to keep the angle shallow.
4.3 Camera Quality
- Resolution: At least 1280×720 px (720p). Higher resolutions (1080p, 4K) can improve accuracy but require more processing power.
- Frame Rate: Ideally 25–30 FPS so the AI engine can track subtle movements and quickly identify new faces.
- Autofocus: Disable it or choose a camera without autofocus. Constant refocusing can cause blurry frames and disrupt face detection.
- Noise-Free Sensor: Low-quality sensors in poor lighting can produce grainy video, leading to more “missed” faces or lower confidence levels.
5. How This Sets the Stage for Installation
This overview provides the foundational knowledge needed before physically installing and configuring your cameras:
- You now know why face detection requires good lighting and angles.
- You understand the workflow from real-time detection on the local device to analytics in the CMS.
- You have a sense of the technical baseline for cameras (HD resolution, stable focus, 2–5 second dwell times).
In the next articles, we’ll dive deeper into:
- Article 2: Good & Bad Cases – Real examples of camera placement pitfalls and how to avoid them.
- Article 3: Installation & Setup Checklist – Step-by-step instructions to ensure proper hardware and software configuration, plus final testing tips.
6. Final Takeaways
- Audience Detection is about measuring real interaction with a display or kiosk.
- Clear, stable video is paramount—grainy, dark, or obstructed views reduce accuracy.
- Angles and distance matter; typical guidelines are 1.5–2.5 meters high, within ~4 meters distance.
- Data privacy is maintained by using numerical face descriptors rather than storing personal imagery.
With this understanding of the basic principles, you’re ready to plan your camera installations and ensure your environment supports accurate face recognition. By following these core guidelines, you’ll set yourself up for success as you move on to the practical examples and step-by-step instructions in the next articles.
Have Questions or Need More Details?
Feel free to reach out to your DISPL technical support contact or consult our additional resources. We’re here to help you make the most of our AI-driven audience detection.
Now you can go to Article 2: Good & Bad Cases, Common Pitfalls
Comments
0 comments
Please sign in to leave a comment.