Customer Behavior Intelligence

This project implements a computer vision pipeline for analyzing customer movement and engagement in retail environments. Developed during my summer school at KAUST, the system goes beyond simple counting by using spatio-temporal filtering to distinguish between passive passers-by and actively engaged customers.

🛠️ Advanced Pipeline Architecture

The system utilizes a multi-stage approach to ensure identity persistence even during occlusions:

• Object Detection: Leverages YOLOv8l for high-accuracy person detection.
• Primary Tracking: Uses DeepSORT with an OSNet (ReID) backbone to handle frame-to-frame associations.
• Post-Tracker Re-Association: A custom PostTracker class manages ID switches and "lost" tracks using a weighted cost function of Cosine Similarity (appearance) and Euclidean Distance (spatial proximity).
• Spatial Logic: Defines multiple Polygonal ROIs (Regions of Interest) to monitor specific store sections.
• Dwell-Time Analytics: Calculates the precise time spent within a zone, filtering out any interactions below a specific temporal threshold (e.g., 3.5 seconds).

📊 Key Features

• Robust Re-Identification: Optimized to maintain consistent IDs for customers even if they temporarily leave a camera's field of view.
• Dynamic Heatmapping: Generates both live decaying heatmaps and cumulative "Master" heatmaps to visualize store-wide traffic patterns.
• Automated Reporting: Finalizes analytics by exporting dwell-time logs for each ROI to CSV files for business intelligence integration.
• Flexible Zone Configuration: Support for multiple concurrent zones (e.g., Track_1, Track_2) with independent entry/exit logic.

💻 Tech Stack

• Core: Python, Jupyter Notebook
• CV Frameworks: ultralytics (YOLO), deep-sort-realtime, torchreid
• Math & Stats: NumPy, SciPy (Cosine distance), OpenCV
• Hardware Acceleration: CUDA-enabled for real-time processing

🎓 Acknowledgments

This project was developed during the KAUST AI Summer School 2025 under the supervision of Dr. Muhammad Mubashar. It serves as a comprehensive customer behavior study for a local retail partner, utilizing advanced computer vision to derive spatial insights.

Project Team:

• Hassan Mohammed Nasr – Project Lead
• Abderrahmene Mehenni
• Khalid Alkaabi