A Generative Latent Space Approach for Real-Time Road Surveillance in Smart Cities

Abstract

Smart cities endeavor to deliver safe and sustainable infrastructure services that enable individuals, organizations, and communities alike to be productive, healthy, informed, and actively involved in rapid urbanization. The widespread installation of closed-circuit television cameras and continuously generated video streams are a strategic data source that can contribute toward safety and sustainability through efficient surveillance of smart city assets and resources. Recent advances in deep learning methods are able to detect and localize salient objects in a video stream. However, a number of practical issues remain unaddressed, such as suboptimality, latency, predictive accuracy, and most importantly the contextualization of all detected salient objects for informed decisions that aligns with ethical surveillance. In this article, we propose a Generative Latent Space (GenLS) approach that overcomes these challenges, specifically in road surveillance. We demonstrate an adaptation of this approach for a prominent use-case in road surveillance, License Plate Detection. GenLS was evaluated for accuracy, robustness, computational cost, and cogency, using a state-of-the-art benchmark dataset on road traffic. Results from these experiments and the corresponding ablation study validate GenLS and confirm its suitability for real-time smart city road surveillance.