Detection method of middle, small and dense traffic signs in automatic driving scene

Abstract

The invention discloses a detection method of middle, small and dense traffic signs in an automatic driving scene. The method comprises following steps of (1) acquiring video data shot by a vehicle dashcam, extracting pictures from the vehicle dashcam, marking traffic signs in the pictures to form a traffic sign data set consisting of <images, target frames> pairs; (2) data preprocessing: preprocessing the traffic sign data set; (3) using the shallow layer network VGG16 as a main network of R-FCN object detection frameworks; (4) improving the VGG16 network model, by use of the characteristics of the shallow layer, reducing the declining times of a characteristic graph and training the RPN network to extract candidate frames; and (5) improving the VGG16 network model, carrying out characteristic combination on the same group of the characteristics of the shallow layer, inputting the combined characteristics into a following R-FCN detection framework, carrying out classification and frame regression on the candidate frames and finally detecting all the traffic signals in the pictures. According to the invention, a detection problem of traffic signs in the automatic driving scene is solved.

Description

technical field [0001] The present invention relates to the technical fields of feature extraction, feature layering and object detection, and in particular relates to a small and dense traffic sign detection method in an automatic driving scene. Background technique [0002] In recent years, autonomous driving has been a hot research and application in the field of artificial intelligence. In autonomous driving scenarios, the detection and recognition of traffic signs is crucial to the understanding of the driving environment. Accurate detection plays a decisive role in subsequent identification, assisted positioning and navigation. For example, through traffic sign detection and identification of speed limit signs to control the speed of the current vehicle; embedding traffic signs into high-precision maps plays a key auxiliary role in positioning and navigation. There are many types of traffic signs with different sizes and angles, so it is difficult to accurately detec...

AI Technical Summary

Problems solved by technology

Another typical representative of the regression-based object detection framework is YOLO and SSD. They do not need to extract candidate frames first, and then classify and adjust the candidate frames. Instead, they directly divide the image into grids. In each grid The corresponding position returns the target position and category information. The whole process of their network training is end-to-end, and the detection speed is very fast, which can fully meet the real-time requirements, but their detection accuracy needs to be improved, especially for the position of small objects. detection

Images