Fast Traffic Sign Recognition with a Rotation Invariant Binary Pattern Based Feature

Shouyi Yin, Peng Ouyang, Leibo Liu, Yike Guo, Shaojun WeiSensors 2015 · OpenReview

TL;DR

This paper introduces a novel "Rotation Invariant Binary Pattern" (RIBP) feature for faster and more robust traffic sign detection, combined with an Artificial Neural Network (ANN) for efficient classification. The method achieves state-of-the-art recognition accuracy while significantly improving processing speed, making it more suitable for real-time driving assistance systems.

What Problem Does This Paper Address?

Safe driving assistance systems heavily rely on accurate and rapid traffic sign recognition. However, real-world driving conditions present significant challenges: traffic signs can appear at various scales, orientations, and under different lighting conditions. Existing methods often struggle to balance robustness against these variations with the speed required for real-time applications. Many robust features are computationally expensive, while faster methods lack the necessary invariance. This paper aims to bridge this gap by developing a feature that is both rotation-invariant and computationally efficient, coupled with an optimized classification approach.

Methods

The proposed method consists of three main stages:

  1. Coarse-grained Location: A standard Hough transform is used to initially identify candidate regions for traffic signs within an image. This step helps to focus subsequent processing on relevant areas, reducing overall computation.
  2. Rotation Invariant Binary Pattern (RIBP) Feature: This is the core novelty. The RIBP feature is designed to be invariant to scale, rotation, and illumination changes. It leverages techniques from ASIFT (Affine Scale Invariant Feature Transform) and combines them with a binary pattern approach.
  3. Affine and Gaussian Space: The method uses affine simulation and Gaussian pyramids to achieve scale and affine invariance.
  4. Binary Pattern: It employs a binary pattern mode to capture local intensity relationships, which is known to be robust to illumination changes.
  5. Circular Histogram: A rotation-invariant circular histogram is constructed around detected "extreme points" (similar to SIFT's extrema). This histogram captures the distribution of binary patterns in a rotationally invariant manner.
  6. Feature Vector: The final feature vector is a 64-dimension representation, significantly smaller than ASIFT's 128 dimensions, contributing to computational efficiency.
  7. ANN-based Dimension Reduction and Classification: To further speed up recognition, an Artificial Neural Network (ANN) is used for two purposes:
  8. Feature Clustering: K-means clustering is employed to group similar feature vectors. An ANN is trained to optimize this clustering process, significantly reducing clustering time.
  9. Classification: The clustered features are then fed into an ANN classifier, which is trained for the specific traffic sign classes. This approach transforms complex feature matching into simpler parameter computations.

Datasets

The paper evaluates its method on two public datasets:

  • German Traffic Sign Recognition Benchmark (GTSRB): Contains over 51,000 German traffic signs across 43 classes, divided into six subsets. Image sizes vary from 15x15 to 250x250 pixels.
  • Sweden Traffic Signs (STS): Includes 20,000 images with 3488 traffic signs, covering various types like pedestrian crossing, speed limits, and parking restrictions. Image sizes range from 3x5 to 263x248 pixels, with an overall image size of 1280x960.

Additionally, the authors collected their own dataset of 25 video clips from real traffic conditions in Beijing, containing 200 Chinese traffic signs under various lighting and weather conditions (night, daytime, raining, sunny).

Key Insights and Ideas

  • Bridging Robustness and Speed: The central idea is that achieving both high robustness (scale, rotation, illumination invariance) and fast processing is possible by carefully designing features and optimizing the classification pipeline.
  • Binary Patterns for Illumination Robustness: The paper highlights the effectiveness of binary patterns in handling illumination variations, a common challenge in real-world vision tasks.
  • Circular Histograms for Rotation Invariance: Using circular histograms around key points is a clever way to encode local information in a rotationally invariant manner, complementing the affine and scale invariance achieved earlier.
  • ANN for Optimizing Clustering: Training an ANN to optimize the K-means clustering process is an interesting approach to reduce the computational burden of feature preprocessing, which is often a bottleneck.
  • Feature Dimension Reduction: The proposed 64-dimension feature vector is significantly smaller than traditional methods like ASIFT, directly contributing to faster processing without a substantial loss in discrimination.

Key Results

  • Processing Speed: The proposed RIBP feature, when combined with ANN classification, achieves significant speed improvements.
  • Compared to ASIFT and ASURF, the feature computation and matching time is reduced by an average of 27.2x and 12.6x, respectively.
  • In real-time traffic sign recognition, the proposed method achieves 43 frames per second (fps) on an embedded platform, a 3.3x improvement over Tang [22].
  • Training time is also reduced, by 61% compared to Tang [22].
  • Recognition Accuracy: The method achieves high recognition accuracy:
  • 98.62% on the GTSRB dataset.
  • 98.33% on the STS dataset.
  • These accuracies are comparable to or better than state-of-the-art methods.
  • Robustness: The method demonstrates robustness in real traffic conditions, correctly detecting 195 out of 200 traffic signs in their collected dataset.
  • Feature Comparison: The proposed feature achieves higher discrimination performance than ASURF and comparable performance to ASIFT, while being significantly faster.

Limitations and Future Work

  • Viewpoint Changes: The paper acknowledges that while the method reduces affine spaces for traffic signs, it creates limitations for applications with very large viewpoint changes, leading to a small recognition accuracy loss.
  • Future Work: The authors suggest enhancing the rotation invariant binary based feature in the affine space to achieve higher robustness to viewpoint changes with a small computation cost.

Summary updated

Abstract

Robust and fast traffic sign recognition is very important but difficult for safe driving assistance systems. This study addresses fast and robust traffic sign recognition to enhance driving safety. The proposed method includes three stages. First, a typical Hough transformation is adopted to implement coarse-grained location of the candidate regions of traffic signs. Second, a RIBP (Rotation Invariant Binary Pattern) based feature in the affine and Gaussian space is proposed to reduce the time of traffic sign detection and achieve robust traffic sign detection in terms of scale, rotation, and illumination. Third, the techniques of ANN (Artificial Neutral Network) based feature dimension reduction and classification are designed to reduce the traffic sign recognition time. Compared with the current work, the experimental results in the public datasets show that this work achieves robustness in traffic sign recognition with comparable recognition accuracy and faster processing speed, including training speed and recognition speed.
BibTeX
@inproceedings{yin2015fast,
  author = {Shouyi Yin and Peng Ouyang and Leibo Liu and Yike Guo and Shaojun Wei},
  title = {{Fast Traffic Sign Recognition with a Rotation Invariant Binary Pattern Based Feature}},
  booktitle = {Sensors},
  year = {2015},
  url = {https://openreview.net/forum?id=av3TToqPMQ},
}
Figure