Scale-Aware Bilateral Feature Pyramid Networks for Traffic Sign Detection

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 intelli-gent driving scene, the existing detection methods are prone to problems such as false detection and missing de-tection. 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 tradi-tional 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 mul-ti-scale foreground target separation. Finally, the lightweight and non-lightweight backbone convolutional net-works are introduced to improve the efficiency and accuracy of the model. The experimental results on the traf-fic 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.