End-to-end lane line detection method and system based on sparse interaction and application thereof

Abstract

The application belongs to the field of artificial intelligence and detection technology, and provides an end-to-end lane line detection method and system based on sparse interaction, which comprises the following steps: extracting multi-scale features of an input road scene image; screening high-confidence lane line region features from the multi-scale features through a binary mask to form a sparse lane line token set; using an attention mechanism to perform sparse interaction and feature updating to obtain enhanced lane line features, wherein the sparse interaction excludes dense interaction with non-lane line region features; initializing a lane line perception query and dynamically predicting adaptive sampling points along the lane line and their weights to extract fine-grained lane line features, wherein the lane line perception query and its content query are iteratively updated in cooperation with each other in terms of position and features; directly predicting and outputting parameters of each lane line to complete end-to-end lane line detection. The application is a complete end-to-end detection method, which can eliminate NMS post-processing and improve small-lane-line detection capability.

Images