摘要:高精度的实时交通标志检测和识别对安全自动驾驶和智能交通系统至关重要。本研究对基线网络YOLOV4进行了升级，增加了多尺度融合模块和注意力机制模块(AMM),丰富了不同尺度交通标志的特征表示。同时,颈部网络结合了特征选择模块和特征对齐模块,增强了高,低层特征图之问像素偏移的语义判别。具体地说,针对AMM,设计了一种转置的自注意力操作其使用互协方差矩阵将令牌维度上的操作转换为通道维度,将时问复杂度从0(㎡?)降低到0(n)。在TT100K交通标志数据集上的实验结果表明,与基线网络(mAPG0.5=76.4%)相比,升级后的网络(mAP@0.5=83.4%)取得了较好的改进,检测和识别速度可达 39.45幢/秒,达到了目前最先进的水平。

Abstraet: Real-time trafie sign detection and recognition with high accuracy is crucial for safe autonomousdriving and intelligent transportation systems. Many deep networks have been proposed, while there is stillroom for further improvement. In this study, the baseline network YOLOV4 was upgraded with a muli-seale fusion module and an attention mechanism module (AMM) for enriching the feature representation oftrafie signs of difierent sizes. eanwhile, a neck network combined a feature selection module and afeature alienment module for enhancing the semantic discriminant of pixel shift between high- and low.level feature maps. $pecifically, a transposed self-attention operation was designed for AMM. lt used cross.covariance matrix for transforming the operation on token dimension to channel dimension and reduced thetime complexity from 0(n?) to O(n). Experimental results on the TT100K traffic sign dataset indicate thatcompared to the baseline network (mAP@0.5=76.4%), the upgraded network achieves good improvement(mAP@0.5 =83.4%) with detection and recognition speed of 39.45 frames per second, and it alsooutperfors several state-of-the-art works.