基于多尺度像素特征融合的实时小交通标志检测算法

交通标志检测技术是先进驾驶辅助系统中重要组成部分。真实的驾驶环境中要求交通标志检测系统具备极高的实时性与准确性。轻量级网络MobileNetv2-SSD能够满足检测的实时性,但准确性不足以满足实际需求。本文将MobileNetv2-SSD作为基础网络,提出一种基于像素重排的多尺度像素特征融合方法,并在网络的检测层引入高效通道注意力机制,实现特征增强。在保证算法的实时性的同时,有效提升了小交通标志的检测性能。实验结果表明,本文算法模型能够在真实环境下准确实时地检测小交通标志。在长沙理工大学中国交通标志检测数据集CCTSDB上取得93.2%的mAP,模型大小仅为17.3M,检测每张图像的时间为0.022 s。 

Real-Time Small Traffic Sign Detection Algorithm Based on Multi-Scale Pixel Feature Fusion

Traffic sign detection technology is an essential part of the advanced driving assistance system. The real-life driving environment requires the traffic sign detection system to have an extremely high real-time performance and accuracy. Lightweight network MobileNetv2-SSD can satisfy real-time detection tasks, but the accuracy can not satisfy the actual requirement. This paper takes MobileNetv2-SSD as the underlying network, proposed a multi-scale pixel feature fusion method based on pixel shuffle, and introduced an efficient channel attention mechanism at the network's detection layer to achieve feature enhancement. The proposed method effectively improves the detection performance of small traffic signs while ensuring real-time performance. Experimental results show that the algorithm model in this paper can detect traffic signs in real environment accurately and in real-time. On the CSUST Chinese traffic sign detection benchmark (CCTSDB), our model obtained 93.2% mAP with the only 17.3M model size, and 0.022 seconds for detecting each image.