Homepage / Case Studies / Identifying Shoplifting Events in Real Time

Identifying Shoplifting Events in Real Time

Disclaimer: Aya Data holds to the confidentiality agreements made with our clients. We shall not disclose any confidential information regarding clients and/or projects, pursuant to the individual agreements made. Consequently, client and project names and any identifiable information may be kept anonymous in our Case Studies.

Overview

There are over 200 million instances of shoplifting per year in the United States alone, which is more than 500,000 a day. This ‘victimless crime’ is neither harmless nor without impact as consumers face higher prices and police and courts struggle to keep up with the burgeoning problem.

With over $120 billion worth of goods shoplifted per year globally, many companies are looking for innovative ways to curb their losses with the use of technology. A leading CCTV analytics provider was building a model to detect instances of shoplifting. They needed accurate labels of shoplifting events from start to finish.

Data Annotation

Image segmentation

Industry

Retail

Headquarters

London, UK

Company Size

250+

Challenge

Aya Data went through 5,000 hours of video surveillance footage and labelled each person in every frame with tight bounding boxes. Once people had been labelled correctly, a ‘key point’ was added to denote the start and end of the theft.

With the subjects and the events having been marked, the data set was used to train a model to analyse shop floors in real time and identify common behaviours indicating a potential theft which could be flagged to staff in the store.

Aya Data’s challenge was to accurately label instances of shoplifting, sometimes with very poor footage and sub-optimal camera angles. The video supplied was from several different stores and was non-sequential, meaning tracking back from a shoplifting event was particularly challenging.

Solutions

Aya Data’s pre-processing team was able to cut 5,000 hours of CCTV footage into manageable slices that were subsequently labelled by over 100 experienced video annotators.

The models demanded the highest levels of quality so a workflow was set up to ensure that each frame was annotated manually, and then checked by a quality assurance lead, before being submitted to the client for review.

Results

The video images labelled by Aya Data were used to train a bespoke computer vision model to detect instances of shoplifting and prior behaviour warnings, with complete satisfaction from the Client.