Evaluation Method of Deep Learning-Based Embedded Systems for Traffic Sign Detection

Miguel Lopez-Montiel,Kenia Picos,Moises Sanchez-Adame,Oscar Humberto Montiel Ross,Ulises Orozco-Rosas

doi:10.1109/access.2021.3097969

Miguel Lopez-Montiel, Kenia Picos + Show 3 more

Open Access

https://doi.org/10.1109/access.2021.3097969

Copy DOI

Abstract

Traffic Sign Detection (TSD) is a complex and fundamental task for developing autonomous vehicles; it is one of the most critical visual perception problems since failing in this task may cause accidents. This task is fundamental in decision-making and involves different internal conditions such as the internal processing system or external conditions such as weather, illumination, and complex backgrounds. At present, several works are focused on the development of algorithms based on deep learning; however, there is no information on a methodology based on descriptive statistical analysis with results from a solid experimental framework, which helps to make decisions to choose the appropriate algorithms and hardware. This work intends to cover that gap. We have implemented some combinations of deep learning models (MobileNet v1 and ResNet50 v1) in a combination of the Single Shot Multibox Detector (SSD) algorithm and the Feature Pyramid Network (FPN) component for TSD in a standardized dataset (LISA), and we have tested it on different hardware architectures (CPU, GPU, TPU, and Embedded System). We propose a methodology and the evaluation method to measure two types of performance. The results show that the use of TPU allows achieving a processing training time 16.3 times faster than GPU and better results in terms of precision detection for one combination.

Highlights

Nowadays, there has been a rapid emergence in the development of Deep Learning (DL) algorithms focused on vision problems for autonomous vehicles
The hyperparameters used for both System 1 (S1) and System 2 (S2) are a batch size of 32, with 2,600 steps, an Adaptive Momentum Estimation (ADAM) optimizer, an initial learning rate of 0.000999, a decay factor of 0.950, and a refresh every 26 steps (1 epoch)
We focused on developing an evaluation methodology based on statistical analysis and algorithms for Deep Learning to be implemented into embedded systems applied to the Traffic Sign Detection (TSD) task for autonomous vehicles

Summary

Introduction

There has been a rapid emergence in the development of Deep Learning (DL) algorithms focused on vision problems for autonomous vehicles. These algorithms have been evolving throughout the years, having different applications such as robotics, object recognition, self-driving cars, among others [1]. Within the applications of autonomous vehicles, different tasks exist to attain vehicle autonomy. One of the main tasks to achieve in computer vision is Traffic Sign Detection (TSD). The vision problem of traffic sign detection started decades ago with some research that focused on classical computer vision algorithms [2], [3], where image processing was used

Methods

Results

Conclusion