AI & Image Processing
Computer Vision Built for Crowded Places
Digeiz models detect, track and classify visitors across hundreds of cameras — even in dense crowds, occlusions and changing light.
Computer vision for complex physical environments
Analyzing visitor behavior in real environments requires computer vision models capable of handling challenging visual conditions and transforming them into reliable behavioral analytics. The Digeiz AI models are designed specifically for these environments and optimized for large venues such as shopping malls, retail stores and transportation hubs.
Typical environments include:
- Dense pedestrian traffic
- Occlusions between individuals
- Overlapping trajectories
- Changing lighting conditions
- Complex indoor layouts
The platform converts raw video streams into structured behavioral analytics through a multi-stage processing pipeline. The key stages include: detection and segmentation, trajectory tracking, counting and flow analytics, demographic classification, and visitor journey reconstruction. Each stage contributes to generating reliable insights on visitor behavior.
Stage 1
Detection and segmentation
The first stage of the AI pipeline detects individuals within each frame of the video stream. Deep neural networks identify human silhouettes and extract visual features used for further analysis.
This stage generates:
- Individual detections
- Spatial coordinates in the image
- Segmentation from the background
These detections are the foundation for trajectory analysis.
Deployment process
Stage 2
Tracking trajectories
After detection, the platform reconstructs the movement of each individual across successive frames. Tracking algorithms connect detections over time to build coherent trajectories.
This process combines:
- Motion prediction
- Appearance similarity
- Spatial consistency
Trajectory reconstruction allows the system to understand how visitors move through the venue.
Stage 3
Privacy-preserving re-identification
To reconstruct visitor journeys across multiple locations, the platform uses appearance-based re-identification. Instead of relying on biometric information, the system generates feature vectors based on visual attributes such as clothing color, body shape and movement patterns.
These feature vectors allow the system to associate observations belonging to the same individual across cameras while preserving privacy. No facial recognition is used and the system cannot identify or authenticate individuals.
Privacy and compliance
Stage 4
Counting and flow analytics
From reconstructed trajectories, the system generates counting and flow metrics. Counting relies on calibrated zones within the video frame. When trajectories intersect these zones, the system can classify events such as entry, exit or pass-by. This mechanism enables accurate measurement of traffic at key points of interest.
Retail store analytics
Stage 5
Demographic segmentation
The platform can extract demographic insights using dedicated neural networks. For each detected trajectory, classification models estimate gender and age group — classified into four brackets: 0-15, 15-25, 25-40, and 40+. Predictions generated across multiple frames are aggregated to produce a final classification for the entire trajectory.
Gender and age group classification accuracy has been independently tested by CESP as part of a 2024 audit. Results showed 96% agreement on gender classification and 85% agreement on age group classification (0-15, 15-25, 25-40, 40+) between Digeiz predictions and manual annotations — with predicted demographic distributions closely matching observed distributions across four shopping centres.
Stage 6
From trajectories to business insights
Once trajectories are reconstructed, the platform aggregates them to generate behavioral insights. These include visitor journeys, dwell times, cross-visits between locations and exposure to retail media screens. By clustering trajectories across the environment, the system builds a structured representation of visitor behavior.
Accuracy and methodology
The performance of the AI models is continuously measured through controlled validation processes. The platform generates analytics across several categories: visitor counting, demographic segmentation, cross-visits, dwell time and exposure to digital media screens.
Accuracy levels vary depending on the metric type, but core counting metrics typically exceed very high reliability thresholds in real-world environments. Analytics are generated for a wide range of points of interest including entrances, stores, corridors, digital screens, kiosks and event areas.
Continuous improvement and production architecture
Computer vision technologies evolve extremely quickly. To ensure the highest level of accuracy, the Digeiz AI team continuously evaluates new model architectures and training approaches.
The platform relies on a fully integrated AI training and evaluation pipeline, allowing the team to rapidly benchmark and deploy improved models at scale. The most recent advances in computer vision are continuously integrated.
Digeiz AI processes video streams real-time locally on servers installed within the client IT — reducing network bandwidth and guaranteeing strong privacy.
Explore the Digeiz platform
Discover how the Digeiz platform transforms camera infrastructures into reliable audience intelligence systems.