Data Augmentation for Occlusion-Robust Traffic Sign Recognition Using Deep Learning

Dineley, Andrew and Natalia, Friska and Sudirman, Sud (2024) Data Augmentation for Occlusion-Robust Traffic Sign Recognition Using Deep Learning. ICIC Express Letters, 15 (4). pp. 381-388. ISSN 2185-2766

Text Friska Natalia_DATA AUGMENTATION FOR OCCLUSION-ROBUST TRAFFIC SIGN RECOGNITION USING DEEP LEARNING.pdf Download (2MB)

Abstract

Traffic sign recognition is an essential feature for self-driving cars. It pro- vides input to the decision-making process when maneuvering through traffic in real time. Correct identification and classification of traffic signs are a challenge because they may be occluded by natural entities, such as leaves and trees, or man-made such as graffiti. In this paper, we present the result of our study into achieving occlusion-robust traffic sign recognition by augmenting the data used to train deep learning models. The data augmen- tation is performed by applying random occlusion of varying coverage percentages to the traffic sign images. We investigated the performance of four different deep network archi- tectures to recognize 11 German speed limit signs using transfer learning techniques on their respective pre-trained models (AlexNet, VGG19, ResNet50, and GoogLeNet). The results of our experiment show that our data augmentation technique improves the recog- nition accuracy at higher occlusion band (61%-70% occlusion) by 17% using GoogLeNet with a slight 2% hit in accuracy at lower occlusion band (1%-10% occlusion). Our study concludes that our data augmentation technique could significantly improve the recogni- tion performance of all models when the traffic sign images are severely occluded.

Item Type:	Article
Keywords:	Traffic sign recognition, Data augmentation, Deep learning, Occlusion, Computer vision
Subjects:	000 Computer Science, Information and General Works > 000 Computer Science, Knowledge and Systems > 006 Special Computer Methods > 006.8 Augmented Reality, Virtual Reality
Divisions:	Faculty of Engineering & Informatics > Information System
Depositing User:	Administrator UMN Library
Date Deposited:	01 Aug 2024 04:28
Last Modified:	01 Aug 2024 04:28
URI:	https://kc.umn.ac.id/id/eprint/31161

Actions (login required)

View Item