Fig 1.
Network structure of ETSR-YOLO.
Fig 2.
Structure of coordinate attention.
Fig 3.
Network structure of ConvNeXt and ConvNeXt Block.
Fig 4.
Composition of the depthwise separable convolution.
Fig 5.
Block designs for the original C3 module, the C3CA module, and the CNeB module.
(a) is the structure of the original C3. (b) is the structure of the C3CA. (c) is the structure of the CNeB.
Fig 6.
Enhanced path aggregation network.
Table 1.
The hardware and software environment used for the experiment.
Table 2.
Model training parameters.
Fig 7.
Visualisation of different categories of traffic signs.
Fig 8.
Statistical visualisation of the TT100K and CCTSDB2021 datasets.
Table 3.
Ablation experiments.
Fig 9.
Precision-recall curve for improved models.
Table 4.
Performance comparison with popular detection models on the TT100K dataset.
Fig 10.
Average precision of YOLOv3, YOLOv5, YOLOv6, YOLOv7 and ETSR-YOLO in each category.
Fig 11.
Confusion matrix of ETSR-YOLO.
Table 5.
Performance comparison of the improved network models on the TT100K dataset.
Table 6.
Performance comparison with popular detection models on the CCTSDB2021 dataset.
Table 7.
Performance comparison of the improved network models on the CCTSDB2021 dataset.
Fig 12.
Sample test results using YOLOv5s and ETSR-YOLO in different scenarios.
(a) is the test results of YOLOv5s. (b) is the test results of ETSR-YOLO.
Fig 13.
External experimental platform.
Table 8.
Comparison of inference speed between YOLOv5s and ETSR-YOLO on the Jetson AGX Xavier platform.
Fig 14.
Detection results on sample sequence using ETSR-YOLO.