Object detection is an essential capability for performing complex tasks in robotic applications. Today, deep learning (DL) approaches are the basis of state-of-the-art solutions in computer vision, where they provide very high accuracy albeit with high computational costs. Due to the physical limitations of robotic platforms, embedded devices are not as powerful as desktop computers, and adjustments have to be made to deep learning models before transferring them to robotic applications. This work benchmarks deep learning object detection models in embedded devices. Furthermore, some hardware selection guidelines are included, together with a description of the most relevant features of the two boards selected for this benchmark. Embedded electronic devices integrate a powerful AI co-processor to accelerate DL applications. To take advantage of these co-processors, models must be converted to a specific embedded runtime format. Five quantization levels applied to a collection of DL models are considered; two of them allow the execution of models in the embedded general-purpose CPU and are used as the baseline to assess the improvements obtained when running the same models with the three remaining quantization levels in the AI co-processors. The benchmark procedure is explained in detail, and a comprehensive analysis of the collected data is presented. Finally, the feasibility and challenges of the implementation of embedded object detection applications are discussed.

Advances in Artificial intelligence (AI) and embedded systems have resulted on a recent increase in use of image processing applications for smart cities’ safety. This enables a cost-adequate scale of automated video surveillance, increasing the data available and releasing human intervention. At the same time, although deep learning is a very intensive task in terms of computing resources, hardware and software improvements have emerged, allowing embedded systems to implement sophisticated machine learning algorithms at the edge. Additionally, new lightweight open-source middleware for constrained resource devices, such as EdgeX Foundry, have appeared to facilitate the collection and processing of data at sensor level, with communication capabilities to exchange data with a cloud enterprise application. The objective of this work is to show and describe the development of two Edge Smart Camera Systems for safety of Smart cities within S4AllCities H2020 project. Hence, the work presents hardware and software modules developed within the project, including a custom hardware platform specifically developed for the deployment of deep learning models based on the I.MX8 Plus from NXP, which considerably reduces processing and inference times; a custom Video Analytics Edge Computing (VAEC) system deployed on a commercial NVIDIA Jetson TX2 platform, which provides high level results on person detection processes; and an edge computing framework for the management of those two edge devices, namely Distributed Edge Computing framework, DECIoT. To verify the utility and functionality of the systems, extended experiments were performed. The results highlight their potential to provide enhanced situational awareness and demonstrate the suitability for edge machine vision applications for safety in smart cities.

Automotive Driver Assistance Systems (ADAS) applications are currently an intensive field of study and innovation. The development of an ADAS is a multidisciplinary task involving electronic hardware design, advanced software implementation, safety considerations and many more. Building an ADAS application implies some challenges that are addressed in this paper. Firstly, all ADAS applications run on highly specific hardware devices embedded in the car with limited computation resources. In this work a novel embedded platform, iADASys, is developed and tested. The platform integrates the elements required to implement an artificial vision based ADAS application such as high performance processor with Deep Learning (DL) computation co-processors or multi-channel high resolution video streaming hardware. Secondly, this work implements an artificial vision application for traffic light detection based on deep neural networks. The model selected in this work is SSD_Mobilenet_V1 and it was trained using Bosch Small Traffic Light (BSTL) dataset. To fulfill real time requirement, the model image input resolution was maintained low at 300x300 pixel. However, the small object size in the dataset together with low resolution lead to poor detection performance. This situation was addressed by fine tuning the model training hyperparameters related to detection scales and aspect ratios. Lastly, the model is deployed in the hardware platform and its performance is measured. Model inference is executed on a specialized mathematical co-processor obtaining the required real time response. The object detection performance is also measured, obtaining promising results.