Traffic Sign Detection Research And Its Application ?

The traffic sign is a kind of important facility in the transportation system and playing a role in many areas in the intelligent driving system.In perception,it is necessary to accurately recognize the category of the traffic sign to assist the driving decision-making.In map and localization,it is necessary to stably detect the instance segmentation of traffic signs.In this paper,the detection and segmentation algorithms of traffic signs are proposed considering the perceptual demand of the intelligent driving system,and a vehicle self-positioning algorithm which using traffic signs is proposed.For the traffic sign detection task,this paper proposes a fine-grained hierarchical architecture object detection network.Traffic signs usually have hundreds of fine-grained categories,many of which are very similar in shape and color.However,traditional object detection networks often recognize fine-grained categories inaccurately and are prone to misclassification.In this paper,the fine-grained categories are aggregated into several superclasses to solve this problem.The super-class is detected firstly.The detection results share the feature layer of the super-class network for finegrain classification.The labeled RoIPooling operation is proposed,thus enables an end-to-end fine-grained detection network.Experiments show that the proposed method achieves good accuracy and can recognize the fine-grained categories accurately.For the traffic sign segmentation task,this paper proposes an instance segmentation method based on edge extraction.The instance segmentation network often requires a large amount of annotation data and occupies a lot of computing resources.However,the traffic signs have definite geometric shapes,thus the traffic signs' segmentation can be obtained by directly extracting the edge.Based on the object detection results,this paper first extracts the candidate edge line by Hough transformation,then uses the DBSCAN clustering algorithm and spatial position filtering to obtain the geometric edge.An edge correction algorithm which based on image gradient is proposed to obtain the refined segmentation result.Experiments show that the proposed method can obtain stable and accurate results in various scenarios.Finally,this paper proposes a vehicle self-positioning method based on traffic signs.Traditional visual positioning algorithms are often affected by dynamic objects and occlusions,whereas traffic signs can solve these problems well.This paper proposed a visual positioning method based on the structure-from-motion framework,and proposed to use traffic sign contour points for multi-frame matching,thus avoiding the errors introduced by feature point mismatching.Finally,the visual positioning results are fused with the IMU and the odometer by extended Kalman filtering.The real-world experiments show that the proposed method can obtain the submeter positioning accuracy and has good robustness.