Detection And Recognition Of Arrow Road Marking Based On L-junction Feature

With the development of intelligent vehicle, autonomous driving technology has been widely concerned. The perception of the surrounding environment and the road information is the key part of the automatic driving technology. Arrow road markings convey important information to autonomous driving system, and there is a great significance to realize their detection and recognition. But detecting and recognizing them is a tough task because they suffer from themselves’ abrasion and non-arrow markings interference, etc. The thesis proposed effective detection and recognition of arrow road markings.We used L-junction string to describe the arrow road markings based on L-junction structure, which makes the effects of interference factors transfer to the deviation on L-junction. To make the detecting model robust enough to deviation in arrow road markings, three encoding methods are proposed to encode those L-junctions within a range as the same code. At the same time, the detecting model sets conditions on detection by using angle feature and topological relation of L-junction on arrow road marking, and reduces the interference of non-arrow marking. In order to determine whether the detection is arrow road markings, we propose an weighted Edit distance algorithm which assign different deviation with different weight to achieve the similarity between detection encode string and standard arrow road marking encoded string, so that the detecting model can achieve higher precision.To test our detect model, three different image sets are collected on the five rings highway of Beijing:clean/dirty arrow road marking images, a video dataset. And then the images of these image sets are detected individually by the detecting model, also calculate the detected accuracy and recall rate. Another deep learning framework (Boosting Convolutional Deep Neural Network (CDNN)) is also implemented for comparison. Extensive experimental results well demonstrate the superior performance of our detecting model.