Data collection and video annotation: weapon detection on the streets

The client’s task was to create a system for recognizing weapons on people in the streets of a large city. To achieve this, it was necessary to create the required data domain and perform weapon detection in public areas and streets. There were no open datasets with weapons, and web scraping provided an insufficient amount of data. The client didn’t know how to organize the data collection process, so they turned to Training Data

To create the necessary data domain, we formed several groups of extras, provided them with replica weapons of various sizes, and mapped routes in the city so that the extras would frequently appear on the external surveillance cameras. Each extra participated in shooting 10 scenes: 5 indoors and 5 on the streets. Subsequently, videos from the cameras were collected and provided to a team of assessors for conducting bounding box detection

• The client received a high-quality dataset according to the technical specifications and trained a neural network with a 99% weapon recognition rate
• From scratch, within a tight timeline, the process of collecting staged videos and photos was organized, and a methodology was outlined for similar tasks