Traffic Light Recognition in Autonomous Driving: Comparison of YOLOv7 and CenterNet2 with HSV Post-Processing

Advanced deep learning networks, such as the YOLO series have been widely utilized for traffic light recognition tasks. However, existing models, overlook the issues of additional noise caused by the external environment(e.g., signal lights of the vehicles on the road, lights on the side of the road...

Full description

Saved in:
Bibliographic Details
Main Authors: Kim, Dayoung, Li, Xingyou, Miraliev, Shokhrukh, Abdigapporov, Shakhboz, Kim, Hyoungrae, Kim, Hakil
Format: Conference Proceeding
Language:English
Subjects:
Automation
autonomous driving
Control systems
Deep learning
HSV post-processing
Predictive models
Reliability
Roads
Task analysis
Traffic light recognition
Online Access:Request full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Advanced deep learning networks, such as the YOLO series have been widely utilized for traffic light recognition tasks. However, existing models, overlook the issues of additional noise caused by the external environment(e.g., signal lights of the vehicles on the road, lights on the side of the road etc.) in the real-world. This study aims to fill this research gap by applying hue-saturation-value(HSV) post-processing technique to the two of the most popular traffic light recognition models YOLOv7 and CenterNet2 and compare the results before and after the usage of the applied technique. After evaluations, the implemented post-processing method showed effectiveness in correctly classifying the detected traffic light signals, resulting in improved overall mAP results from 82.6% to 90.1 % for the YOLOv7 model and from 83.5% to 91.8% for the CenterNet2 model.
ISSN:2642-3901
DOI:10.23919/ICCAS59377.2023.10317045