Research on a Traffic Sign Recognition Method under Small Sample Conditions

Xiao Zhang, Zhenyu ZhangSensors 2023 · DOI

TL;DR

This paper tackles the challenge of accurately recognizing traffic signs when training data is scarce, a common issue due to rapid sign updates and high labeling costs. They propose an improved Few-Shot Object Detection (FSOD) framework that enhances feature extraction, refines candidate region proposals with an attention mechanism, and effectively fuses multi-scale features, leading to better performance on small-sample traffic sign recognition tasks.

What Problem Does This Paper Address?

The core problem addressed is the difficulty in training robust traffic sign recognition systems due to the limited availability of labeled data. Traffic signs are constantly updated, and acquiring and annotating large datasets is resource-intensive. Traditional object detection methods rely on vast amounts of data, making them unsuitable for scenarios with few training samples. This scarcity leads to issues like overfitting and the generation of irrelevant candidate boxes, hindering detection accuracy. The paper specifically aims to improve traffic sign recognition under these "few-shot" conditions.

Methods

The proposed method builds upon the Few-Shot Object Detection (FSOD) framework and introduces several key improvements:

  • Adjusted Backbone Network: The authors modify the ResNet backbone (specifically reducing it to 44 layers from 50) to better capture image details and semantic information, while also reducing computation. They also introduce dropout in the fully connected layer to mitigate overfitting.
  • ACBAM-RPN Module: A novel Region Proposal Network (RPN) module, termed ACBAM-RPN, is introduced. This module integrates a multi-attention mechanism (based on CBAM - Convolutional Block Attention Module) to enhance feature representation. The attention mechanism helps the RPN focus on relevant candidate boxes and suppress irrelevant ones, improving the accuracy of target detection. The CBAM itself comprises channel and spatial attention modules.
  • Multi-Scale Feature Fusion (FPN): A Feature Pyramid Network (FPN) is incorporated to effectively extract and fuse features across multiple scales. This is crucial for handling variations in traffic sign sizes, especially in few-shot scenarios where scale differences between support and query sets can be significant. The FPN helps create a feature pyramid that captures both high-resolution, detailed features and low-resolution, semantic features.

Datasets

The paper utilizes two main datasets:

  • Self-Made Dataset: A custom dataset was created by crawling traffic sign images and annotating them using LabelImg. This dataset contains seven categories of traffic signs (straight, nohook, stop, left, slow, crosswalk, right). Data augmentation techniques like rotation, clipping, and scaling were applied to expand this dataset and mitigate overfitting.
  • PASCAL VOC Dataset: This standard dataset was used for comparative experiments. For few-shot tasks, specific subsets of categories were selected as "new classes" in different splits (Split1, Split2, Split3) to simulate few-shot scenarios.

Key Insights and Ideas

  • Attention for Candidate Refinement: The core idea of using an attention mechanism within the RPN is to intelligently filter candidate regions. Instead of relying on generic proposals, the model learns to prioritize regions that are more likely to contain traffic signs, directly addressing the problem of irrelevant candidate boxes in low-data regimes.
  • Multi-Scale Fusion for Scale Invariance: The integration of FPN is critical for handling the inherent scale variations in traffic sign images. By fusing features from different levels of the network, the model gains a more comprehensive understanding of objects at various sizes, which is particularly challenging in few-shot learning.
  • Backbone Modification for Detail and Regularization: Adjusting the backbone network not only aims to improve feature extraction but also to combat overfitting. Reducing network depth and adding dropout are standard regularization techniques that prove effective in small-sample settings.

Key Results

  • Self-Made Dataset Performance: The proposed ACBAM-FSOD method achieved significant improvements on the self-made dataset. For the 5-way 3-shot task, it improved AP50 by 4.27% and AP75 by 1.64% compared to the baseline. For the 5-way 5-shot task, improvements were 1.64% for AP50 and 0.93% for AP75.
  • Comparative Experiments (PASCAL VOC): When compared against other few-shot object detection algorithms (Meta RCNN, TFA, MPSR) on the PASCAL VOC dataset, the proposed method demonstrated competitive or superior performance across various splits and shot numbers, particularly in AP50 and AP75 metrics. For instance, on Split1 (5-way 5-shot), the proposed method achieved 40.4% AP50 and 47.6% AP75, outperforming other methods in most configurations.

Limitations and Future Work

  • Real-time Performance: The authors acknowledge that while their method improves accuracy, most current few-shot object detection methods are based on two-stage detectors, which can be slow and may not meet real-time requirements. They suggest that future work should focus on improving the real-time performance of these models.
  • Further Exploration: The paper calls for continued experimental exploration to further enhance the performance of target detection in small-sample scenarios, implying that there is still room for improvement.

Summary updated

Abstract

Traffic signs are updated quickly, and there image acquisition and labeling work requires a lot of manpower and material resources, so it is difficult to provide a large number of training samples for high-precision recognition. Aiming at this problem, a traffic sign recognition method based on FSOD (few-shot object learning) is proposed. This method adjusts the backbone network of the original model and introduces dropout, which improves the detection accuracy and reduces the risk of overfitting. Secondly, an RPN (region proposal network) with improved attention mechanism is proposed to generate more accurate target candidate boxes by selectively enhancing some features. Finally, the FPN (feature pyramid network) is introduced for multi-scale feature extraction, and the feature map with higher semantic information but lower resolution is merged with the feature map with higher resolution but weaker semantic information, which further improves the detection accuracy. Compared with the baseline model, the improved algorithm improves the 5-way 3-shot and 5-way 5-shot tasks by 4.27% and 1.64%, respectively. We apply the model structure to the PASCAL VOC dataset. The results show that this method is superior to some current few-shot object detection algorithms.
BibTeX
@inproceedings{zhang2023research,
  author = {Xiao Zhang and Zhenyu Zhang},
  title = {{Research on a Traffic Sign Recognition Method under Small Sample Conditions}},
  booktitle = {Sensors},
  year = {2023},
  doi = {10.3390/s23115091},
}
Figure