Traffic Sign Recognition Based on Up-sampling Convolution

Yitian Lu, Ping Jiang, Shun Nishide, Xin Kang, Fuji Ren, N Dalal, B Triggs, Avd S Nchez, R Girshick, J Donahue, T Darrell, S Ren, K He, G Gkioxari, Doll R P, J Redmon, S Divvala, T Y Lin, P Goyal, H Zhao, J Shi, X Qi, J Long, E Shelhamer, A Shrivastava, A Gupta, S Zhang, L Wen, X Bian, Z Zhang, S Qiao, C Xie, L C Chen, M Collins, Y Zhu, G Lin, A Milan, C Shen, G Papandreou, I Kokkinos, J Stallkamp, M Schlipsing, J Salmen, C Igel, Z Shou, J Chan, A Zareian, Yuming Xu, X Wang, C Jin, W Liu, X Sun, M Lv, C QuanProceedings of 2019 the 9th International Workshop on Computer Science and Engineering 2019 · DOI

Abstract

This paper presented a method makes traffic sign recognition faster and more accurate.

Introduction

Detecting traffic signs is a critical yet challenging task for autonomous driving systems, primarily due to the "small object" detection problem. Standard Convolutional Neural Networks (CNNs) often struggle with a structural trade-off: high-level layers possess strong semantic understanding but lose geometric detail, while low-level layers retain resolution but lack contextual intelligence. While multi-scale detectors like the Single Shot Multibox Detector (SSD) attempt to address this, they often fail on very small or blurry objects because their low-level predictions are not semantically rich enough.

Proposed Methodology: Simplified Up-sampling SSD

To bridge the gap between semantic depth and geometric precision, this research introduces a Simplified Up-sampling SSD architecture. The model incorporates several key innovations:

  • Streamlined Feature Extraction: The backbone is simplified compared to standard VGG-based SSDs to maintain processing speed and prevent the loss of small features in excessively deep layers.
  • Deconvolutional Layers: The architecture integrates successive deconvolutional (up-sampling) layers. These layers "stretch" high-level semantic maps back to a higher resolution.
  • Feature Fusion: By up-sampling, the network fuses "smart" deep features with "sharp" early-layer geometric features. This allows for predictions based on feature maps that are both high-resolution and semantically informed.
  • Optimization: The study found that the Adam optimizer provided faster convergence and superior stability compared to standard SGD for this specific architecture.

Experimental Dataset

The model was evaluated using The German Traffic Sign Recognition Benchmark (GTSRB), which consists of over 51,000 images across 43 classes. The dataset presents significant real-world challenges, including motion blur, occlusion, and extreme lighting variations.

Examples of overly bright and dim traffic sign samples from the GTSRB dataset.

Results and Discussion

The proposed up-sampling approach significantly outperformed baseline SSD models. The modified architecture achieved a 97.6% Mean Average Precision (mAP), a substantial leap from SSD300 (79.7%) and SSD512 (73.4%).

The Depth Paradox

A notable finding was that the deeper SSD512 performed worse than the shallower SSD300. This suggests that for small object detection, excessive down-sampling in deep networks is counterproductive, as it destroys the fine-grained features necessary for recognition. The up-sampling layers effectively "reconstruct" these lost details.

Data Augmentation Insights

Interestingly, standard data augmentation techniques did not consistently improve performance. For specialized, low-resolution datasets like GTSRB, complex augmentation can sometimes introduce noise that hinders the model's ability to learn distinct features, suggesting a need for more tailored augmentation strategies.

Conclusion

By integrating deconvolutional layers into the SSD framework, this research demonstrates a robust method for enhancing small object detection in traffic scenarios. Future work will explore applying this up-sampling logic to other domains requiring high geometric precision, such as human pose estimation and joint detection.

Summary updated

Abstract

This paper presented a method makes traffic sign recognition faster and more accurate.
BibTeX
@inproceedings{lu2019traffic,
  author = {Yitian Lu and Ping Jiang and Shun Nishide and Xin Kang and Fuji Ren and N Dalal and B Triggs and Avd S Nchez and R Girshick and J Donahue and T Darrell and S Ren and K He and R Girshick and K He and G Gkioxari and Doll R P and J Redmon and S Divvala and R Girshick and T Y Lin and P Goyal and R Girshick and H Zhao and J Shi and X Qi and J Long and E Shelhamer and T Darrell and A Shrivastava and A Gupta and R Girshick and S Zhang and L Wen and X Bian and Z Zhang and S Qiao and C Xie and L C Chen and M Collins and Y Zhu and G Lin and A Milan and C Shen and L C Chen and G Papandreou and I Kokkinos and J Stallkamp and M Schlipsing and J Salmen and C Igel and Z Shou and J Chan and A Zareian and Yuming Xu and Fuji Ren and Xin Kang and Shun Nishide and Ping Jiang and X Wang and C Jin and W Liu and X Sun and M Lv and C Quan and F Ren},
  title = {{Traffic Sign Recognition Based on Up-sampling Convolution}},
  booktitle = {Proceedings of 2019 the 9th International Workshop on Computer Science and Engineering},
  year = {2019},
  doi = {10.18178/wcse.2019.06.020},
}
Figure