Lane line detection method and system

Abstract

The invention discloses a lane line detection method and system, and the method specifically comprises the steps: carrying out the modeling of a lane detection task, building a convolutional neural network model, and carrying out the feature extraction of a convolutional feature graph; in the training stage, collecting lane line training sample images, increasing the diversity of training samples through preprocessing, and further obtaining convergent network model parameters through iterative training; in a reasoning stage, post-processing a reasoning result to obtain a lane line predicted by the model; according to the invention, the mechanism is introduced into the network structure, the semantic features of the extracted feature map are richer through the fusion of local and global information, and the positioning precision of the far end of the lane line is further improved by using the network model trained by the structure.

Description

technical field [0001] The invention relates to a lane line detection technology, in particular to a lane line detection method and system. Background technique [0002] Lane lines, as an important part of road markings, can effectively guide intelligent vehicles to drive within the constrained road structure area. Real-time detection of road lane lines is an important link in the intelligent vehicle assisted driving system, which helps to assist in path planning , Carry out functions such as road deviation warning, and can provide reference objects for positioning and navigation. [0003] At present, the most advanced lane line detection methods in the industry are based on CNN. For example, SCNN and SAD networks regard lane line detection as a semantic segmentation task, which has a heavy encoding and decoding structure. However, this method usually uses a small image as input, which makes it difficult to accurately predict the far end of the curved lane line; in addition...

AI Technical Summary

Problems solved by technology

[0003] At present, the most advanced lane line detection methods in the industry are based on CNN. For example, SCNN and SAD networks regard lane line detection as a semantic segmentation task, which has a heavy encoding and decoding structure. However, this method usually uses a small image as input, which makes it difficult to accurately predict the far end of the curved lane line; in addition, the above methods are usually limited to the detection of a predefined number of lane lines, for the lane line detection task on the actual road, in the image The scenes are different, and the number of lane lines is not fixed; for this PointLaneNet follows the strategy based on the candidate area, by generating multiple candidate lines in the image, thus getting rid of the inefficient encoder and the number of predefined lanes limits

[0004] The existing ResNet122 is used as the backbone network to extract semantic features, the input image is passed through the backbone to obtain the corresponding feature map, and then each grid on the feature map is regarded as an anchor, and the anchor predicts the lane line corresponding to the grid. After that, the NMS algorithm is used to retain the lane lines with high confidence, and filter out the redundant candidate lane lines; and then through the end-to-end training, the output of the entire lane line can be directly obtained, and at the same time, the output of the fixed number of lane lines can be got rid of However, using the anchor to predict the entire lane line passing through its corresponding grid on the obtained feature map makes the performance drop severely when predicting the far end of the curve; due to the limitation of the convolution kernel scale, the feature map obtained by the above scheme Only local area information can be captured, and long-distance and short-distance context information cannot be captured at the same time

Images