AI Retail Theft Detection

Inspiration

Retail theft costs stores billions every year, and small and mid-sized retailers often can't afford full-time security or someone watching cameras all day. Staff can't monitor every aisle, and wrongly accusing a customer hurts trust. We wanted to build an AI tool that helps stores notice likely theft early while keeping a human in charge of every decision.

What it does

ShelfGuard AI analyzes store video and flags suspicious behavior patterns, such as concealing items, so staff can review them quickly.

• Detects people and products in the video and tracks their movement
• Flags behavior patterns that may indicate shoplifting
• Sends an alert to staff with the relevant clip
• Focuses on behavior, not identity: no facial recognition and no profiling of individuals
• Every alert is reviewed by a person before any action is taken

How we built it

• Computer vision: we used object detection and pose estimation to track people and their hand and body movements across frames
• Behavior analysis: a machine learning model classifies movement patterns as normal or suspicious
• Pipeline: video frames → detection and tracking → behavior classification → alert to staff
• Tech stack: Python, OpenCV, and PyTorch, with a simple web interface to view alerts
• We tested the system on sample retail-style video and tuned it to balance catching real events against false alarms

Challenges we ran into

• Limited data: real shoplifting footage is rare and sensitive, so we relied on available datasets and simulated scenarios
• False positives: normal actions, like putting an item in your own bag or basket, can look suspicious, so we tuned the alert threshold and kept a human reviewing every alert
• Privacy and bias: we deliberately avoided facial recognition and built the system around actions rather than who a person is
• Time: building and testing a full pipeline within the hackathon was a tight squeeze

Accomplishments that we're proud of

• Built a working end-to-end prototype, from raw video to a staff alert
• Designed the system with privacy and fairness in mind from the start
• Turned a hard, sensitive problem into a tool that supports staff instead of replacing their judgment

What we learned

• How to build and evaluate a computer vision pipeline, and why data quality matters so much
• That false positives are a big deal in real-world AI, because a wrong alert affects a real person
• That responsible AI design, meaning privacy, human oversight, and transparency, is essential for surveillance-adjacent tools

What's next for ShelfGuard AI

• Test on more diverse, real-world store environments
• Add explainable alerts that show why something was flagged
• Reduce false positives with more data and better tuning
• Integrate with existing store camera systems and alerting tools

Built With

• computer-vision
• deep-learning
• machine-learning
• numpy
• object-detection
• opencv
• python
• pytorch
• yolo