Neural Network Lane Polyline Generation via BEV Features
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current autonomous driving systems face challenges in accurately detecting and generating lane polylines, which is crucial for safe driving, as existing methods often suffer from imperfect geometry transformation and information loss during view transformation processes.
Innovation Solution
The proposed solution involves using a neural network model that includes a first neural network for extracting bird's eye view features and a second neural network for generating lane polylines, with specific loss functions such as translation loss, embedding offset loss, seed probability loss, and order loss to minimize errors and improve accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If view transformation is used to generate lane polylines, then lane detection can be performed, but geometry transformation errors and information loss occur
Solution Approach 1:
The patent transforms the 2D image space into a 3D point cloud space and then projects it back to 2D for polyline generation. This dimensional transformation approach avoids direct view transformation of the original 2D image, preserving geometric information while enabling accurate lane detection through multi-dimensional feature extraction and projection.
2Reliability
If traditional view transformation methods are used, then lane detection is achieved, but geometry transformation errors occur
Solution Approach 1:
The patent introduces a 3D point cloud as an intermediary representation between the 2D input image and the final lane polyline output. By converting image features to 3D points and then projecting them, the system maintains geometric accuracy while enabling reliable lane detection, avoiding direct transformation errors.
3Measurement precision
If multiple sensors are used to improve detection accuracy, then lane detection precision improves, but system complexity increases
Solution Approach 1:
The patent merges data from multiple sensors (camera, LiDAR, radar) into a unified 3D point cloud representation. By combining these different sensor types into a single coordinated coordinate system and processing them together through the neural network, the system achieves high detection precision while managing complexity through integrated processing.
Data Source
AI summary
The present disclosure relates to a method and apparatus for generating a lane polyline by using a neural network model. The method according to an embodiment may extract a multi-scale image feature by using a base image obtained from at least one sensor loaded in a vehicle. According to the method, the multi-scale image feature is input to a first neural network model as input data and a BEV feature may be obtained as output data from the first neural network model. Also, according to the method, the BEV feature may be input to a second neural network model as input data and a polyline image with respect to a certain road may be obtained as output data from the second neural network model. In the present disclosure, a lane polyline obtained from the neural network may be utilized in vehicle control without going through an additional treatment process.