ROAD INFORMATION DETECTION BASED ON IMPROVED MASK RCNN ALGORITHM OF FPT

In view of the poor effect of target detection algorithm applied to road information target detection and the problems of missed detection and false detection, an improved mask RCNN road information detection algorithm based on FPT (feature pyramid transformer) is proposed to help urban road construction to improve the quality of work. In the feature fusion network, FPT based on transformer idea was introduced to replace the original FPN to fuse features across space and scale, so as to achieve the effect of feature enhancement and improve the accuracy of model detection. In the experiment, the idea of transfer learning was used to pre-train PASCAL-VOC2012 data set, and the pre-training weight was obtained. The experimental results show that the average accuracy of the algorithm is improved by 7.5/10.6 percentage points compared with the original algorithm when ResNet50/ResNet101 is used respectively, and the performance on small targets is better than other commonly-used algorithms.