- Case study ↓
- AI/ML
- Computer Vision
Theft Detection That Reviews Hours of CCTV For You
AI that analyses store CCTV footage, scores suspicious customer behaviour and timestamps likely theft, so security teams go straight to the moments that matter instead of watching hours of footage.
The demo
See Theft Detection at work
- Industry
- Computer Vision
- Input
- Store CCTV footage
- Output
- Confidence scores, Timestamps, FastAPI
- Built with
- YOLOv8 · DeepSORT · SlowFast · CLIP · FastAPI
The problem
Nobody can watch every camera all the time
Shoplifting and shrinkage cost retailers heavily, yet most CCTV only records. Security staff can't watch dozens of screens at once, and the moment of concealment lasts only seconds.
Going back through footage after a loss is slow and easy to get wrong. Detection that relies only on a product disappearing misses how theft actually happens.
The solution
What Theft Detection does
Behaviour analysisLooks at how people act, not just whether an item went missing
Confidence scoresEvery alert is ranked, so staff check the likeliest first
Long-video processingHours of footage reviewed automatically
Works on existing CCTVNo new hardware needed to start
Timestamped resultsGo straight to the moment that matters
Technology
How it’s built
Every step, from the store CCTV footage to the results.
Step 1
Input
- CCTV recordings
Step 2
Detection & tracking
- YOLOv8
- DeepSORT
- SAM2
Step 3
Action recognition
- SlowFast
- ResNet3D
- CLIP
Step 4
Output
- Confidence scores
- Timestamps
- FastAPI
- AI approaches tested to find the most reliable one
- 8
- CCTV recordings processed in a single run
- 1–3 h
Who it’s for
- RetailersReducing shrinkage without adding more security staff.
- Loss-prevention teamsWho want to focus on real incidents instead of watching screens.
- Security and CCTV companiesAdding behaviour analytics to their own product through an API.
Case study
Theft Detection: the full case study
- Overview
The Retail Theft Detection AI System is a video analytics solution that helps retailers find suspicious customer behaviour in existing CCTV footage. Instead of security staff watching screens all day, the system analyses movement and activity and highlights possible theft for review.
- Problem
Shoplifting and inventory shrinkage cause major losses. Watching dozens of cameras continuously is slow and error-prone, and ordinary CCTV only records, so security teams miss incidents or spend hours reviewing footage.
- Solution
ML Bench built a video analysis platform that reviews CCTV footage and flags behaviour linked to theft. Rather than relying only on products going missing, it looks at customer actions, movement patterns, interactions with merchandise and sequences of behaviour, then gives each event a theft confidence score.
Several AI methods were researched and tested, and the final hybrid design combines them to improve accuracy while keeping false alarms low. The architecture is built to extend to real-time monitoring across multiple stores.
- Challenges & solutions
Detecting theft only from objects disappearing produced too many false alarms.Analysed customer behaviour and action sequences instead, using a hybrid of several models.
No single AI approach was clearly best for retail footage.Evaluated 8 approaches on real surveillance footage and chose the most reliable combination.
Store recordings run for hours.Built an offline inference pipeline that processes 1–3 hour videos and timestamps each detection.
- Results
- 8 AI approaches evaluated to find the best theft detection method
- Hybrid architecture gave the most reliable behaviour detection of all approaches tested
- Far fewer false positives than detection based on objects disappearing
- 1–3 hour CCTV videos processed automatically, with timestamps for each event
Have Questions?
Find the right solution
for now
Usually not. The system works with existing CCTV that gives a clear view of the shop floor. We check your camera angles during the free assessment.
Why ML Bench
Why teams choose ML Bench
Built for real-world footageWe build computer vision for real cameras and scans: retail, sports, healthcare, logistics and safety.
Starts with the data you haveExisting cameras, recordings or image archives are usually enough for a first proof of concept.
Custom to your setupWe adapt the model to your camera angles, image quality, classes and workflow, not the other way round.
One team, pilot to productionData, model training, dashboard, API and deployment handled by the same team.
Next AI/ML project →
UV Scan / Hygiene
CV-based kitchen heatmap system that maps high-activity zones from overhead video for hygiene audits and layout planning.
Free assessment
See What AI Spots in Your Store
Send us a short clip from one of your cameras and we'll show you what the model picks up, free.
- UI/UX Design
- Web Development
- Mobile Apps
- SEO
- AI/ML Solutions
