| More Details: |
Propose an optimized and improved traffic sign detection model based on YOLOv8n, addressing the issues of low accuracy and inaccurate detection, especially in adverse weather conditions, observed in current traditional network models. We leverage existing techniques, including the BoTNet (Bottleneck Transformers for Visual Recognition) module to enhance image classification capabilities, the ODConv (Omni-dimensional Dynamic Convolution) module to supplement attention for improved accuracy, and the LSKA (Large Separable Kernel Attention) module to reduce memory and computational complexity while enhancing small object detection capabilities. Additionally, we employ the WIoU (Wise Intersection over Union) loss function to enhance the model’s generalization performance. Without additional preprocessing to simulate adverse weather conditions, our results on the TT100K dataset, including mAP50, mAP95 (mAP is ’mean Average Precision’), and F1, relative to the original YOLOv8n model, show improvements of 3%, 4%, and 2.5% in misty conditions and 1%, 2.6%, and 1.7% in dark conditions, respectively. On the GTSDB dataset, in misty conditions, the improvements are 5%, 4.2%, and 2.3%, and in dark conditions, the improvements are 3%, 6.6%, and 6%. When 30% of the training set is augmented with fog, the detection performance of the improved model is comparable to, or even exceeds, that of the YOLOv8n model trained with the entire fog-augmented training set. This comprehensive result highlights the significant superiority of our model over the comparative model, demonstrating its practical applicability. |
