Customer Demographics from Video Analysis
The Problem
A grocery store chain wanted to understand their customer demographics for marketing purposes — specifically, how many people enter the store, their gender distribution, approximate age ranges, and how many children accompany shoppers. Traditional methods like manual counting or customer surveys are expensive, inaccurate, and don't scale. The client needed an automated solution that could analyze existing security camera footage without requiring new hardware or disrupting store operations. They also needed proof-of-concept results before committing to a full-scale rollout across multiple locations.
The Solution
We developed a two-stage pipeline. First, to avoid burdening the client with footage collection during the testing phase, we used Bicirai (bicir.ai) to generate realistic synthetic videos of people entering a store — providing diverse test data without privacy concerns. For person detection and tracking, we implemented Ultralytics YOLO11n, which reliably identifies and follows individuals frame-by-frame as they cross the entrance threshold.
The gender and age classification proved trickier. Traditional face-based detection models failed — store entrance cameras capture people at a distance and from angles where faces aren't clearly visible. After testing several approaches, we discovered that OpenCLIP ViT-B-32 could classify gender and age from the full body appearance rather than requiring a clear face. The model analyzes clothing style, body shape, and overall posture to distinguish between adult men, adult women, boys, and girls. This body-based approach proved far more reliable than face detection for real-world entrance footage. Results are aggregated into a structured database with timestamps, enabling analysis by hour, day, or week.
The Result
The experiment achieved approximately 90% accuracy across gender detection, age estimation, and child identification. The system successfully generated marketing-ready spreadsheets showing peak shopping times, demographic breakdowns, and family shopping patterns. This proof-of-concept gave the client concrete data to present to upper management, demonstrating the feasibility of full-scale deployment. The synthetic video approach also proved valuable — it allowed rapid iteration and testing without waiting for real footage, significantly accelerating the development timeline.