Building a high-quality annotated image library to improve object detection

Modern cities worldwide are facing considerable pressures to improve city attractiveness to dwellers and businesses. A key factor to raise the liveability of cities is the existence of a good network of walkways and cycleways. To implement such infrastructures under growing resource constraints, cities need to know beforehand how people are using public spaces. One of the best ways to study people behaviour in cities is to make use of CCTV systems with the help of computer vision models. This project aims to collect a set of CCTV-like images from four city centres and annotate persons, cyclists and vehicles with the purpose of: developing an object detection model to be deployed on CCTV cameras; and building a shareable repository of annotated images for use by other developers. The collection of images were captured between January and February 2022 with a high definition camera similar to the type used in CCTV systems. The annotation work started in January 2022 and ended in March 2022. The annotation comprised of a team of five annotators and two reviewers working with specialized software. The project annotated 99,246 unique objects in 10,446 images. The most annotated object class was “Pedestrian” with 81.9% of the total number of annotated objects.