A novel traffic sign recognition approach for open scenarios
Traffic sign recognition systems based on the traditional deep learning technologies typically follow the complete data-driven mode, resulting in their unstable performances and significant security risks when applied to the real-world open scenarios. To alleviate this problem, a novel method is pro...
Saved in:
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Science Press (China Science Publishing & Media Ltd.)
2023-05-01
|
| Series: | Shenzhen Daxue xuebao. Ligong ban |
| Subjects: | |
| Online Access: | https://journal.szu.edu.cn/en/#/digest?ArticleID=2512 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Traffic sign recognition systems based on the traditional deep learning technologies typically follow the complete data-driven mode, resulting in their unstable performances and significant security risks when applied to the real-world open scenarios. To alleviate this problem, a novel method is proposed by constructing the semantic data set based on road traffic sign design standards and using the zero-shot learning (ZSL) mechanism to develop a general TSR framework with reasoning and interpretation capabilities. This method can effectively overcome the problems of dynamic update of road traffic signs and classes missing in practice. Furthermore, the national standard for road traffic signs is used to abstract the general attributes of all classes and then the information is injected into the training process of traditional data-driven model as domain knowledge. With the help of domain knowledge, the proposed ZSL-based TSR method can recognize traffic signs that have not been seen in the training stage more accurately than random prediction and traditional deep learning models. Experimental results on the Chinese traffic sign database (CTSDB) and the German traffic sign recognition benchmark (GTSRB) demonstrate that our method, which trains a semantic auto-encoder model, can significantly improve the accuracy in traditional zero-shot learning settings. Specifically, when identifying previously unseen traffic signs in the training set, our approach achieves an improvement in accuracy of at least 29.96% and 24.25% on CTSDB and GTSRB, respectively, compared to random prediction. The study verifies the feasibility and effectiveness of the proposed scheme. |
|---|---|
| ISSN: | 1000-2618 |