Lane Boundary Detection via Segment Vector Clustering

The patent segments the lane boundary detection task into multiple components: extracting road boundary candidates, classifying them into different types (solid lines, dashed lines, curved sections), and processing them through separate machine learning models. This segmentation allows the system to handle the broad range of lane boundary variations systematically rather than attempting to handle all variations simultaneously.

The patent changes parameters such as converting 2D image segments into segment vectors, adjusting classification thresholds, and modifying model training parameters to adapt to different lane boundary appearances. These parameter changes enable the machine learning models to generalize across various lane boundary types without requiring manual tagging of every possible variation.

The patent implements self-service through automated road boundary candidate extraction and classification. The system automatically generates training data by processing images through computer vision algorithms that identify and categorize lane boundaries without requiring manual annotation. This self-service approach produces sufficient training data while eliminating the costly manual tagging process.

The patent creates copies of road boundary information in the form of segment vectors and classified boundary representations. These copied data structures serve as training data that capture the essential features of lane boundaries without requiring original manually tagged images. The copying process preserves the necessary information for model training while avoiding the cost of manual annotation.

The patent performs preliminary actions by automatically extracting and classifying road boundary candidates before training the machine learning models. The system pre-processes images to identify potential lane boundaries, categorize them by type and characteristics, and prepare training data in advance. This preliminary action reduces the time needed for data collection and processing while ensuring the training data is ready for model training.

The patent maintains continuous useful action through an automated pipeline that continuously processes images, extracts road boundary candidates, classifies them, and prepares training data without interruption. The system operates continuously to accumulate diverse training data representing various lane boundary types, ensuring sufficient data volume for model generalization while minimizing time loss through uninterrupted processing.