Global lane line detection method based on key points and gradient equalization loss

Abstract

The invention discloses a global lane line detection method based on key points and gradient equalization loss, and the method comprises the steps: S1, carrying out collection, preprocessing and laneline marking of an image, and forming a lane line data set in a multi-driving scene; S2, designing a group of key point global representation lane lines, and obtaining refined key point sequence labels by adopting a spline interpolation algorithm; S3, constructing a lane line key point detection model, and adopting a gradient equalization loss function optimization model for training; and S4, during detection in a model in a real vehicle, inputting an image frame and outputting a lane line classification probability graph and a key point regression vector graph to acquire a final lane line detection result through key point matching combination and non-maximum suppression post-processing. A full-convolution deep learning method is adopted, through feature key point detection and gradientequalization training, the method can adapt to lane line detection in a large-range complex driving scene, and high robustness, real-time performance and positioning precision are achieved.

Description

technical field [0001] The invention belongs to the field of computer vision and vehicle intelligent driving, and more specifically relates to a global lane line detection method based on key points and gradient equalization loss. Background technique [0002] Lane detection is one of the basic tasks and key technologies in the field of vehicle intelligent driving. Intelligent driving technologies such as lane departure warning, lane keeping, and vehicle autonomous navigation can only be realized if the vehicle perceives its position in the lane and provides a stable and reliable data source for subsequent path planning and decision-making control. In practical applications, lane line detection requires high algorithm robustness and real-time performance. In appearance, the lane lines have solid / dashed lines, yellow / white lines, etc., and often appear blurred and damaged to varying degrees. In terms of driving conditions, the vehicle needs to adapt to various complex envir...

AI Technical Summary

Problems solved by technology

[0004] However, the method based on convolutional neural network mainly uses bounding box detection and image semantic segmentation to realize lane location, and there are still some limitations: 1) The lane is a slender linear object, and the bounding box location will introduce a lot of interference information, reducing the Detection accuracy; 2) The pixel-level detection network has a large amount of calculation and poor real-time performance, and the post-clustering processing ignores the global information of the lane line, which is prone to false detection; 3) The global detection of the lane line is often easy to lose the local details of the target. low precision

Images