面向交通标志检测的尺度感知双向特征金字塔网络

Scale-Aware Bilateral Feature Pyramid Networks for Traffic Sign Detection

实时精准的交通标志检测是实现自动驾驶和智能交通的重要技术之一.针对真实智能驾驶场景中背景复杂且交通标志尺度较小,现有的检测方法容易出现错检和漏检等问题,提出一种尺度感知的双向特征金字塔网络,实现复杂交通场景中实时、精准的交通标志检测.首先,为解决微小标志在传统金字塔网络中尺度丢失的问题,通过构建自底向上和自顶向下的双向金字塔网络,循环地学习尺度感知的融合特征;然后引入前景注意力模块和尺度感知损失函数,学习和优化不同尺度下的前景显著特征和关联,实现多尺度前景目标分离;最后,引入轻量级和非轻量级主干卷积网络,可以同时提高模型效率和精度.在真实复杂场景的交通标志数据集TT100K和STSD中的实验结果表明,该方法的检测精度达到了66.7%和60.9%,同时实时检测速率达到了30帧/s.

Real time and accurate traffic sign detection is one of the important technologies to realize automatic driving and intelligent transportation.Due to the complex background and small scale of traffic signs in the real intelligent driving scene,the existing detection methods are prone to problems such as false detection and missing detection.Therefore,a scale-aware bilateral feature pyramid networks for real-time accurate traffic sign detection(SbFPN)in complex traffic scenes is proposed.Firstly,to solve the problem of scale loss of small signs in traditional pyramid network,SbFPN builds bottom-up and top-down two-way pyramid network,cyclically calculate scale-aware feature fusion.Then,the foreground attention module and scale-aware loss function are introduced to learn and optimize the foreground salient features and correlation at different scales,so as to realize multi-scale foreground target separation.Finally,the lightweight and non-lightweight backbone convolutional networks are introduced to improve the efficiency and accuracy of the model.The experimental results on the traffic sign data sets TT100K and STSD in real complex scenes show that the method achieves the best accuracy of 66.7% and 60.9%,and the real-time detection rate reaches 30 frames per second.