Towards large-scale, automated, accurate detection of CCTV camera objects using computer vision. Applications and implications for privacy, safety, and cybersecurity. (Preprint)

TL;DR

This paper introduces CCTVCV, the first computer vision models capable of accurately detecting CCTV cameras in images and videos, supporting privacy and safety routing in mapping systems.

Contribution

We developed and evaluated the first MS COCO-compatible CCTV camera detection models, achieving up to 98.7% accuracy, filling a critical gap in privacy-aware mapping solutions.

Findings

CCTVCV models achieve up to 98.7% accuracy in CCTV detection.

Built on 8,387 annotated images with 10,419 CCTV instances.

Comprehensive comparison and analysis of detection models and challenges.

Abstract

In order to withstand the ever-increasing invasion of privacy by CCTV cameras and technologies, on par CCTV-aware solutions must exist that provide privacy, safety, and cybersecurity features. We argue that a first important step towards such CCTV-aware solutions must be a mapping system (e.g., Google Maps, OpenStreetMap) that provides both privacy and safety routing and navigation options. However, this in turn requires that the mapping system contains updated information on CCTV cameras' exact geo-location, coverage area, and possibly other meta-data (e.g., resolution, facial recognition features, operator). Such information is however missing from current mapping systems, and there are several ways to fix this. One solution is to perform CCTV camera detection on geo-location tagged images, e.g., street view imagery on various platforms, user images publicly posted in image sharing platforms such as Flickr. Unfortunately, to the best of our knowledge, there are no computer vision models for CCTV camera object detection as well as no mapping system that supports privacy and safety routing options. To close these gaps, with this paper we introduce CCTVCV -- the first and only computer vision MS COCO-compatible models that are able to accurately detect CCTV and video surveillance cameras in images and video frames. To this end, our best detectors were built using 8387 images that were manually reviewed and annotated to contain 10419 CCTV camera instances, and achieve an accuracy of up to 98.7%. Moreover, we build and evaluate multiple models, present a comprehensive comparison of their performance, and outline core challenges associated with such research.