In smart city scenarios, the huge proliferation of monitoring cameras scattered in public spaces has posed many challenges to network and processing infrastructure. A few dozen cameras are enough to saturate the city’s backbone. In addition, most smart city applications require a real-time response from the system in charge of processing such large-scale video streams. Finding a missing person using facial recognition technology is one of these applications that require immediate action on the place where that person is. In this paper, we tackle these challenges presenting a distributed system for video analytics designed to leverage edge computing capabilities. Our approach encompasses architecture, methods, and algorithms for: (i) dividing the burdensome processing of large-scale video streams into various machine learning tasks; and (ii) deploying these tasks as a workflow of data processing in edge devices equipped with hardware accelerators for neural networks. We also propose the reuse of nodes running tasks shared by multiple applications, e.g., facial recognition, thus improving the system’s processing throughput. Simulations showed that, with our algorithm to distribute the workload, the time to process a workflow is about 33% faster than a naive approach.
Tiny computers called System-on-a-Chip like Raspberry Pi have revolutionized the development of applications for Smart Home and Smart City. Some Machine Learning algorithms have been used to process a large amount of data produced by these Internet of Things (IoT) devices. An important issue in the context of processing IoT data is the decision on where the machine learning algorithm will run. To support this decision, it is necessary to classify the IoT devices according to their capabilities to run these algorithms, in terms of CPU performance, required memory, and energy demand. The aim of this paper is to classify IoT devices according to their capabilities to run machine learning algorithms, and reporting real experiments that validate the proposed classification.
9 February 2023 in Journal of the Brazilian Computer Society
Smart city platforms provide several services to facilitate the development of applications. Such platforms typically manage several applications, deal with a large volume of data, and serve many devices and users that generate a high volume of requests. The large number of requests to handle and the complex operations to perform often cause overloads on the platform, degrading the quality of service provided to users and applications. In this context, monitoring the underlying computational infrastructure in which smart city platforms and applications are deployed and the platform operations is essential. The monitoring process can allow for examining fluctuations in the behavior of the platform's components to detect performance degradation and overloads (including unforeseen ones), contribute to avoiding interruptions in the platform's services, and increase its scalability to assimilate significant amounts of requests, devices, and users. This paper presents a strategy and architecture to enable the non-intrusive monitoring of operations on smart city platforms and their underlying infrastructure. The proposal covers monitoring at multiple levels and is based on the aspect-oriented programming (AOP) paradigm so that it is possible to monitor the platform's operations without intervening in the platform's implementation or generating coupling regarding monitoring. This paper presents the implementation of the monitoring architecture and its instantiation in the context of Smart Geo Layers (SGeoL), a platform that has been used in several real-world smart city applications. This paper also reports the results of computational experiments to evaluate the performance of the proposed monitoring architecture for response time to requests, CPU usage, and RAM utilization. The obtained results show an evident increase in response time with the number of simultaneous requests and a significant correlation between the response time and the CPU utilization in the deployment of the monitoring architecture.
Intelligent surveillance systems have been increasingly relying on edge-based Artificial Intelligence for real-time object detection and tracking. However, fixed cameras suffer from a limited field of view and degraded image quality at long distances, while resource-constrained edge devices struggle to run complex vision models without incurring high latency and thermal issues. This paper presents CANELA (Camera with Autonomous Navigation and Envisioning Led by AI), a low-cost pan-tilt edge camera system for active object tracking. CANELA combines a custom gimbal, parallel control algorithms, and a Raspberry Pi 5 with a Hailo-8L M.2 Entry-Level accelerator to enable efficient on-device inference and dynamic camera repositioning to keep targets centered. Experimental results in a face-detection pipeline leveraging hardware acceleration demonstrate that CANELA maintains high frame rates while preserving CPU availability by offloading neural network inference to the accelerator. Compared to fixed cameras, it increases the duration of high-quality target capture, improving detection performance and tracking stability.