Map-Generation Large Model Training for End-to-End Road Vectorization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current map generation methods for autonomous driving are inefficient and prone to errors due to human intervention or multi-stage processing, leading to uncertain quality and poor accuracy.
Innovation Solution
A training method for a map-generation large model that utilizes a multi-task learning framework with regression, classification, and segmentation modules to perform end-to-end map generation from road top-view samples, adjusting model parameters based on matching results to improve accuracy and avoid error accumulation.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Manufacturing precision
If human-intensive processes and multi-stage processing are used for map generation, then current methods can produce maps, but the process is inefficient and prone to errors with poor accuracy
Solution Approach 1:
The patent merges multiple processing stages (road element detection, classification, and vectorization) into a single integrated large model that processes road top-view images end-to-end, generating complete map data including vectorized point sets, category labels, and segmentation masks in one operation, thereby improving both accuracy and efficiency
Solution Approach 2:
The large model is designed with multi-functional capabilities, simultaneously performing road element detection, classification into different categories (lane lines, road surfaces, signs), and vectorization into precise geometric representations, replacing multiple specialized tools with a single universal system
2Reliability
If multi-stage processing is used for map generation, then detailed map data can be obtained, but error accumulation occurs leading to uncertainty
Solution Approach 1:
The patent consolidates multiple processing stages into a single integrated large model that processes road top-view images end-to-end, generating complete map data including vectorized point sets, category labels, and segmentation masks in one operation, thereby improving both accuracy and efficiency
Solution Approach 2:
The model employs iterative training with labeled training samples that include ground truth annotations, allowing the system to learn from correct answers and continuously improve its performance, reducing errors through feedback-driven optimization
3Extent of automation
If human-intensive processes are used for map generation, then current methods can produce maps, but reliance on human intervention increases costs and reduces efficiency
Solution Approach 1:
The large model enables autonomous map generation by automatically detecting road elements, classifying them into different categories, and vectorizing them into precise geometric representations without human intervention, allowing the system to serve itself and eliminate manual labor
Solution Approach 2:
The patent replaces human manual annotation and processing with an automated deep learning-based large model that uses neural networks to perform detection, classification, and vectorization tasks, substituting mechanical human labor with intelligent automated systems
Data Source
AI summary
A training method for a map-generation large model is provided, including: obtaining a training sample set, each training sample in the training sample set including a road top-view sample, a first vectorized point set and a first category of a first road element, and a first mask of the road top-view sample; inputting the road top-view sample into an initial map-generation large model, and correspondingly outputting a second vectorized point set and a second category of the second road element, and a second mask of the road top-view sample; determining a model loss according to a matching result between the second and first road element, the first vectorized point set, the first category, the first mask, the second vectorized point set, the second category and the second mask, and adjusting a parameter of the initial map-generation large model according to the model loss to obtain a map-generation large model.