多尺度上下文融合的交通标志识别算法研究

Research on Traffic Sign Recognition Algorithm Based on Multi-scale Context Fusion

针对自然场景中小型交通标志检测漏检和虚警问题,提出了一种多尺度上下文融合的交通标志检测算法。以YOLOv3为检测框架,将特征金字塔网络的深层特征信息融合进更浅层特征层,提高更浅层特征层中高级前景语义信息的利用率;在YOLOv3框架基础上加入上下文模块,重新分配交通标志特征图中的上下文信息权重,加强目标特征信息的复用;使用融合预测目标置信度的网络损失函数来进行端到端的训练。在中国交通标志数据集TT100K上试验75类小型交通标志获得了56.93%的平均精度均值,相比于YOLOv3算法,所提算法精度提高了9.19%,验证了所提算法的有效性,表明了在小目标和多目标的环境下所提算法检测效果提升明显。

To solve the problems of missed detection and false alarm of small and medium-sized traffic signs in natural scenes, a traffic sign detection algorithm based on multi-scale context fusion is proposed, in which YOLOv3 is used as the detection framework.First, the deep layer feature information of the feature pyramid network is fused into the shallower feature layer to improve the utilization rate of the high-level foreground semantic information in the shallower feature layer.Then, a context module is added on the basis of YOLOv3 framework to redistribute the weight of context information in the traffic sign feature image to strengthen the reuse of target feature information.Finally, end-to-end training is carried out using the network loss function fused with the predicted target confidence.An average accuracy of 56.93% is obtained by experiments on TT100 K-75 data set of traffic signs in China, which is 9.19% higher as compared with the YOLOv3 algorithm, verifying the effectiveness of the proposed algorithm and indicating that the detection effect of the proposed algorithm is improved significantly in the environment of small targets and multiple targets.