Road Geometry Estimation Using BEV Transformer Object Tracking
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing road geometry tracking solutions face challenges with inadequate accuracy and high computational and memory requirements, making them less suitable for real-time implementations in autonomous vehicles.
Innovation Solution
A computer-implemented method and system that encodes images from vehicle cameras using machine-learning algorithms to extract image features, transforms these features into a Bird's Eye View representation, and decodes it to extract object embeddings using transformer-based algorithms, thereby improving road geometry estimation accuracy without significant increases in computational power or data storage.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional road geometry tracking methods are used, then computational resources and memory storage are consumed, but accuracy is inadequate
Solution Approach 1:
The patent segments the road geometry tracking problem into multiple discrete road objects (lane markers, road edges, barriers) and processes each object independently through detection, classification, and tracking stages. This segmentation allows the system to manage computational complexity by handling individual objects rather than the entire road scene as a single complex problem.
Solution Approach 2:
The patent transforms 2D image features from camera views into a 3D Bird's Eye View representation, adding a dimensional transformation step. This dimensionality change enables more accurate road geometry estimation by providing a top-down perspective that better represents the actual road layout, while the transformation is optimized to control computational costs.
2Measurement precision
If traditional road geometry tracking methods are used, then computational resources are consumed, but memory storage requirements are high
Solution Approach 1:
The patent extracts only the essential features and parameters needed for road geometry tracking from the full image data. By identifying and extracting key road objects and their geometric properties (positions, orientations, shapes) rather than storing complete image sequences, the system reduces memory storage requirements while maintaining tracking accuracy.
Solution Approach 2:
The patent implements a tracking framework that discards redundant or low-value data while recovering and retaining critical geometric information across time steps. The system selectively maintains only the necessary road object parameters for continuous tracking, eliminating unnecessary data storage while preserving the essential geometric representation needed for accurate road model estimation.
3Measurement precision
If advanced machine learning algorithms are used, then accuracy is improved, but processing time increases
Solution Approach 1:
The patent performs preliminary detection and classification of road objects in advance, organizing the data into structured formats before the tracking phase. By pre-processing images to identify and classify road objects (lane markers, edges, barriers) beforehand, the system reduces the computational burden during real-time tracking, thereby decreasing processing time while maintaining detection accuracy.
Solution Approach 2:
The patent transforms image parameters into a different coordinate system (Bird's Eye View) and represents road objects using optimized geometric parameters. This parameter transformation simplifies the tracking computations by working with directly relevant geometric properties rather than raw pixel data, improving processing efficiency while preserving detection accuracy.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Computer-implemented methods and related aspects for determining a representation of one or more road objects of a road for a vehicle traveling on the road are disclosed, where the vehicle has one or more cameras. The method comprises, for each time step out of a plurality of consecutive time steps, encoding one or more images output from the one or more cameras using one or more machine-learning algorithms trained to output image features of one or more road objects depicted in an image provided as input to the one or more machine-learning algorithms. The method further comprises transforming a plurality of image features comprised in the one or more encoded images to a Bird's Eye View (BEV) representation of the plurality of image features. Moreover, the method comprises decoding the BEV representation in order to extract a set of object embeddings from the BEV representation using one or more transformer-based machine-learning algorithms trained to output the set of object embeddings based on an input comprising the BEV representation, a set of object queries, and a set of transformed prior object embeddings extracted at a preceding time step. Further, the method comprises outputting a position and class of each road object of the one or more road objects by decoding the extracted set of object embeddings.