Is Edge Computing Always Suitable for Image Analysis? An Experimental Analysis

In the era of Smart Cities, video surveillance stands as a pivotal tool for enhancing urban security, optimizing resource management, and improving the quality of urban life. When video surveillance is seamlessly integrated with image analysis systems, raw visual data are transformed into actionable insights, significantly enhancing the capability of Smart Cities to ensure public safety and optimize urban operations. Image analysis systems mainly rely on the cloud: images are offloaded to a cloud infrastructure to be processed, analyzed and segmented for inference. The analysis of images in external systems, however, is not always recommended, due to privacy/security concerns, e.g., human action recognition. In this paper, we investigate the opportunity to adopt edge computing to implement such systems, where images are analyzed directly on-premises. To investigate the suitability of this approach, we carried out an extensive experimentation using two large-scale Fed4Fire+ testbeds, namely, Grid'5000 and Virtual Wall. Specifically, we considered different cloud-edge configurations using different inference models, and evaluated the impact of those models on performance and resource utilization. Based on these results, we provide a set of guidelines for the adoption of different models depending on the requirements of the specific application.