CDFF: a fast and highly accurate method for recognizing traffic signs

Abstract

Convolutional neural networks method is a commonly used traffic sign recognition method based on deep learning over recent years. However, traffic signs contain objects of different sizes. Since small objects occupy a small input image area, the features that can be extracted are less, and the detection difficulty is greater than that of medium and large objects, and it is still challenging to achieve high-speed and high-accuracy detection of all objects of different sizes at the same time. In this paper, a model for detecting traffic signs is proposed, namely CDFF and CDFF-s. The model contains the following four modules: (1) in the backbone part of the model, we apply an improved activation function FMish to increase training stability, (2) after the backbone of the model, we apply the DFb-SPP module to perform context and semantic fusion, (3) in the neck part of the model, we use the DFb module for feature fusion, which also reduces the number of parameters, and (4) in the head part of the model, we propose a loss function SCIoU, which is optimized for small objects and the model is converged faster. The experimental results on the general traffic sign datasets TT100K and LISA show that the proposed two models can achieve accurate small object detection without losing the detection accuracy of medium and large objects. In addition, excellent results are also obtained on the remote sensing dataset RSOD with similar object size distribution. Meanwhile, the detection speed is faster than YOLOv4, which can meet the accuracy and real-time requirements of automatic driving systems and assisted driving systems.

Appendix

1.1 Evaluating metrics

To quantitatively evaluate the proposed approach, we follow previous work [10,11,12,13, 17, 46]. The evaluation metrics include AP (Average Precision) series, AR (Average Recall) series and mAP (mean Average Precision). With Precision as the vertical coordinate and Recall as the horizontal coordinate, the area composed of the PR (Precision-Recall) curve and the coordinate axis is AP. mAP is the average value of AP for each category, which contains both Precision and Recall, and is the main evaluation metric of the model.

The main evaluation metrics are defined as follows:

• FPS: Frames Per Second, how many frames of images can be detected in one second.

• Detecting results:

  TP = True positive

  TN = True negative

  FP = False positive

  FN = False negative

• P: ${\text{Precision}} = \frac{{{\text{TP}}}}{{{\text{TP}} + {\text{FP}}}}$

• R: ${\text{Recall}} = \frac{{{\text{TP}}}}{{{\text{TP}} + {\text{FN}}}}$

• AP: ${\text{AP}} = \mathop \int \limits_{0}^{1} p\left( r \right){\text{d}}r$

• mAP: ${\text{mAP}} = \frac{{\mathop \sum \nolimits_{i = 1}^{K} {\text{AP}}_{i} }}{K}$

• AR: ${\text{AR}} = 2\mathop \int \limits_{0.5}^{1} {\text{recall}}\left( o \right)do$

• AP_S and AR_S: AP and AR for small objects of area smaller than $32^{2}$;

• AP_M and AR_M: AP and AR for objects of area between $32^{2}$ and $96^{2}$;

• AP_L and AR_L: AP and AR for large objects of area bigger than $96^{2}$.

$K$ is the number of categories.