一种基于注意力模型的无锚框交通标志识别算法 
下载： 533次
Anchor-Free Traffic Sign Recognition Algorithm Based on Attention Model
天津大学电气自动化与信息工程学院, 天津 300072
图 & 表
图 1. AAFCNN模型的结构
Fig. 1. Structure of AAFCNN model
下载图片 查看原文
图 2. 语义连接路径的结构
Fig. 2. Structure of semantic connection path
下载图片 查看原文
图 3. 不同的注意力模块。(a)通道注意力模块;(b)空间注意力模块
Fig. 3. Different attention modules. (a) Channel attention module; (b) spatial attention module
下载图片 查看原文
图 4. 注意力模型的结构
Fig. 4. Structure of attention model
下载图片 查看原文
图 5. TT100K数据集中交通标志的尺寸分布
Fig. 5. Size distribution of traffic signs in TT100K dataset
下载图片 查看原文
图 6. 三种尺度交通标志的准确率-召回率曲线。(a)(0,32]的像素区间;(b)(32,96]的像素区间;(c)(96,400]的像素区间
Fig. 6. Accuracy-recall curves of traffic signs at three scales. (a) Pixel interval of (0,32); (b) pixel interval of (32,96]; (c) pixel interval of (96,400]
下载图片 查看原文
图 7. AAFCNN模型的部分可视化识别结果
Fig. 7. Part of visual recognition results of AAFCNN model
下载图片 查看原文
表 1不同交通标志识别方法的性能对比
Table1. Performance comparison of different traffic sign recognition methods
| Method | Backbone | Params /106 | Index | S /% | M /% | L /% |
|---|
| Faster R-CNN | ResNet-101 | 52.2 | Recall | 72.0 | 91.3 | 91.5 | | Precision | 76.1 | 87.5 | 86.1 | | F1-score | 74.0 | 89.4 | 88.7 | | Faster R-CNN +FPN | ResNet-101 | 60.1 | Recall | 86.6 | 95.5 | 95.1 | | Precision | 85.0 | 92.9 | 92.3 | | F1-score | 85.8 | 94.2 | 93.7 | | Ref. [15] | | 81.2 | Recall | 87.4 | 93.6 | 87.7 | | Precision | 81.7 | 90.8 | 90.6 | | F1-score | 84.5 | 92.0 | 89.1 | | RetinaNet | ResNeXt-101 | 94.7 | Recall | 87.4 | 95.1 | 93.1 | | Precision | 84.3 | 95.9 | 94.2 | | F1-score | 85.8 | 95.5 | 93.6 | | FCOS | ResNeXt-101 | 89.7 | Recall | 88.7 | 95.6 | 92.4 | | Precision | 85.6 | 96.4 | 93.5 | | F1-score | 86.8 | 96.0 | 93.0 | | CenterNet | HourglassNet | 191.3 | Recall | 89.7 | 96.0 | 92.4 | | Precision | 90.1 | 96.7 | 94.9 | | F1-score | 89.9 | 96.3 | 93.6 | | AAFCNN | DenseNet-121 | 48.1 | Recall | 90.6 | 95.6 | 93.1 | | Precision | 91.2 | 97.3 | 96.8 | | F1-score | 90.9 | 96.4 | 94.9 |
|
查看原文
表 2密集连接网络的深度对识别性能的影响
Table2. Effect of depth of densely connected network on recognition performance
| Backbone | Params /106 | AP /% |
|---|
| S | M | L |
|---|
| DenseNet-121 | 48.1 | 63.4 | 80.1 | 86.1 | | DenseNet-169 | 65.4 | 62.5 | 79.9 | 86.1 | | DenseNet-201 | 101.4 | 61.7 | 79.7 | 85.7 | | DenseNet-264 | 154.8 | 61.9 | 80.0 | 85.0 |
|
查看原文
表 3注意力模型的位置对识别性能的影响
Table3. Effect of location of attention model on recognition performance
| Location | Params /106 | AP /% |
|---|
| S | M | L |
|---|
| In coding path | 48.1 | 63.4 | 80.1 | 86.1 | | In decoding path | 47.8 | 62.1 | 80.0 | 85.8 | | Both coding path and decoding path | 48.2 | 62.6 | 80.0 | 85.1 |
|
查看原文
表 4各模块的性能对比
Table4. Performance comparison of each module
| Model | Params /106 | AP /% |
|---|
| S | M | L |
|---|
| Base | 14.1 | 60.8 | 79.7 | 85.9 | | Base+AM | 14.2 | 61.7 | 79.8 | 85.1 | | Base+SCP | 47.8 | 61.9 | 80.0 | 87.2 | | Base+AM+SCP | 48.1 | 63.4 | 80.1 | 86.1 |
|
查看原文
褚晶辉, 黄浩, 吕卫. 一种基于注意力模型的无锚框交通标志识别算法[J]. 激光与光电子学进展, 2021, 58(16): 1610020. Jinghui Chu, Hao Huang, Wei Lü. Anchor-Free Traffic Sign Recognition Algorithm Based on Attention Model[J]. Laser & Optoelectronics Progress, 2021, 58(16): 1610020.
PDF全文
