Automatic driving panoramic lane line model training and detection method

• Main Authors: ZHAN DONGXU, FENG XUYANG, SHAN YUMEI
• Format: Patent
• Language: Chinese; English
• Published: 18.10.2022
• Subjects: CALCULATING; COMPUTING; COUNTING; PHYSICS

The invention discloses an automatic driving panoramic lane line model training and detection method, which comprises the following steps of: 1) data acquisition, cleaning and labeling: acquiring multiple segments of video data of multiple visual angles through a camera, and labeling lane lines in a segmented broken line mode; 2) model regression branch training: keeping the original states of front-view and rear-view pictures and directly sending the pictures into the model for training; respectively rotating the left front image, the right left rear image, the right front image and the right rear image anticlockwise or clockwise, then merging the images into one image in the horizontal direction, and feeding the images into a model for training after rotating and merging; 3) model classification branch training: 4) model deployment reasoning, so that the problem that the side small-angle lane lines cannot be detected due to the Line Anchor prior direction can be greatly overcome, and the lane lines can be w