ARCHIVES

The expansion of infrastructure for smart cities relies on creating large amounts of real-time data from an increasing number of cameras, sensors, and Internet of Things (IoT) devices for purposes such as traffic management, surveillance, and environmental monitoring. Although cloud-based deep learning systems possess substantial computational power, they suffer from high latency issues, excessive bandwidth usage, and significant privacy concerns, rendering them unsuitable for mission-critical applications. Edge computing provides solutions to many of these foundational issues through the processing of data collected. However, given that edge devices have limited resources, the standard deep learning models are difficult to use on these devices. This review provides a resourceful overview of the development of lightweight deep learning architectures used in edge environments, including MobileNet, Tiny YOLO, and TinyML-based methods, as well as advanced optimization techniques such as pruning and quantization. The review presents a comparison of data sets, from traffic to surveillance, public object identification, and IoT sensing, concerning the following parameters: Model accuracy, computational efficiency, and feasibility for real-time use. Additionally, this review identifies numerous primary areas requiring future study, particularly because of the limited number of diverse real-world datasets currently available for training and testing algorithms and the lack of validation at the hardware level of edge platforms. This review also describes current trends and provides insight into the development of efficient, scalable, and deployable edge AI solutions for smart city systems.