Document Type
Event
Start Date
28-4-2022 5:00 PM
Description
In this study, we propose a novel method to detect small traffic signs that appeared in dashboard camera images. Our method is a framework consisting of the following three distinct algorithms. Grouping Window, Super Resolution Generative Adversarial Network (SRGAN), and a two-stage cascade classifier. Potential regions of interest (ROI) are extracted with Grouping Window which is a sophisticated modification of the traditional sliding window technique. The ROI are upsampled and enhanced using SRGAN. Then the traffic signs among high-resolution ROI are detected and identified by the two-stage cascade classifier of which the first stage filters the ROI that do not contain traffic signs and the second stage classifies the ROI that contains traffic signs into respective classes. The proposed method is capable of detecting traffic signs in the 5-8 square-pixel range. The detection of small objects in this square-pixel range is not generally addressed by state-of-the-art frameworks such as R-CNN and YOLO. We trained our method by using the German Traffic Sign Recognition Benchmark dataset (GTSRB) and tested it on random dashboard camera images containing small traffic signs. The experimental results on 15 random dashboard camera images show that our baseline model localizes 10 of the 23 small traffic signs belonging to the aforementioned pixel range and produces 70 false positives in total. Also, it classifies only one of the detected traffic signs correctly into 43 classes. We plan to improve our method by using image denoising techniques and comparing results.
Included in
GR-210 - Detection of Small Traffic Signs Using Image Super Resolution
In this study, we propose a novel method to detect small traffic signs that appeared in dashboard camera images. Our method is a framework consisting of the following three distinct algorithms. Grouping Window, Super Resolution Generative Adversarial Network (SRGAN), and a two-stage cascade classifier. Potential regions of interest (ROI) are extracted with Grouping Window which is a sophisticated modification of the traditional sliding window technique. The ROI are upsampled and enhanced using SRGAN. Then the traffic signs among high-resolution ROI are detected and identified by the two-stage cascade classifier of which the first stage filters the ROI that do not contain traffic signs and the second stage classifies the ROI that contains traffic signs into respective classes. The proposed method is capable of detecting traffic signs in the 5-8 square-pixel range. The detection of small objects in this square-pixel range is not generally addressed by state-of-the-art frameworks such as R-CNN and YOLO. We trained our method by using the German Traffic Sign Recognition Benchmark dataset (GTSRB) and tested it on random dashboard camera images containing small traffic signs. The experimental results on 15 random dashboard camera images show that our baseline model localizes 10 of the 23 small traffic signs belonging to the aforementioned pixel range and produces 70 false positives in total. Also, it classifies only one of the detected traffic signs correctly into 43 classes. We plan to improve our method by using image denoising techniques and comparing results.