Evaluation of deep learning algorithms for traffic sign detection to implement on embedded systems

Nowadays, machine learning algorithms are trendy and are used to solve different problems of autonomous vehicles obtaining good results. Among these algorithms, deep learning has emerged as an excellent alternative to improve the results of the state-of-the-art in machine vision applications. An essential task in autonomous vehicles is the detection of traffic signs. Some metrics used for these detectors focus on assessing precision and recall. However, it is necessary to consider other factors, such as the implementation of these models on an embedded system. In this work, we implement deep learning algorithms on an embedded system to evaluate two different detection algorithms: Faster R-CNN and Single Shot Multibox Detector (SSD) with two feature extractors, ResNet V1 101 and MobileNet V1 to determine the location of traffic signs within the observed scenario. The contribution of this work focuses on evaluating the implementation of traffic sign detection systems based on deep learning algorithms on embedded systems. The experiments were achieved on the experimental embedded system board Nvidia Jetson Nano. The inference time and memory consumption of these detection systems were evaluated; they delivered good performance (81–98%) measure by average precision for each superclass (prohibitory, warning, and mandatory).