Attention-Based Lane Detection Using Masked Feature Maps

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional lane recognition techniques in autonomous driving systems require post-processing for filtering non-lane road regions and line-fitting, which can be inefficient and inaccurate.

Innovation Solution

A method using a lane mask to detect lanes in input images through a multi-step process involving convolutional neural networks (CNNs), including generating feature maps, segmentation score maps, and embedded feature maps, followed by lane masking, convolutional operations, and loss value backpropagation to learn parameters for improved lane detection and filtering.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If conventional lane recognition techniques are used, then lane detection can be performed, but post-processing is required for filtering non-lane road regions and line-fitting

Engineering Contradiction:
Improvelane detection accuracyVSAvoidprocessing complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent merges the lane detection function and post-processing functions (filtering and line-fitting) into a single integrated neural network model. The second CNN receives feature maps from the first CNN and simultaneously performs lane candidate detection, filtering of non-lane regions, and line-fitting operations within its architecture, eliminating the need for separate post-processing steps and reducing overall system complexity while maintaining detection accuracy

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The second CNN is designed as a multi-functional module that performs multiple tasks: it detects lane candidates, filters non-lane road regions using the segmentation score map, and generates fitted lane parameters. This universal module consolidates what were previously separate processing functions into one component, reducing the number of processing stages required

Inventive Principle:
Principle #6Universality (Multi-functionality)

2Measurement precision

If post-processing processes are applied for filtering and line-fitting, then lane recognition can be completed, but processing time increases

Engineering Contradiction:
Improvelane parameter accuracyVSAvoidprocessing time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The first CNN performs preliminary feature extraction and generates segmentation score maps that pre-identify potential lane regions before the second CNN processes them. This preliminary action prepares the data in advance, allowing the second CNN to focus only on relevant regions and perform filtering and line-fitting more efficiently, reducing overall processing time while maintaining precision

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent establishes a continuous processing flow where the output of the first CNN feeds directly into the second CNN without intermediate discrete processing steps. The feature maps and segmentation scores are continuously transformed through the neural network layers, enabling parallel computation of filtering and line-fitting operations that maintains processing speed while achieving accurate lane parameters

Inventive Principle:
Principle #20Continuity of useful action

Data Source

PatentEP3686779B1Method and device for attention-based lane detection without post-processing by using lane mask and testing method and testing device using the same
Publication Date: 2024.01.24 STRADVISION
  • EP3686779B1 patent drawingFigure 1
  • EP3686779B1 patent drawingFigure 2
  • EP3686779B1 patent drawingFigure 3A

AI summary

A method of attention-based lane detection without post-processing by using a lane mask is provided. The method includes steps of: a learning device instructing a CNN to acquire a final feature map which has been generated by applying convolution operations to an image, a segmentation score map, and an embedded feature map which have been generated by using the final feature map; instructing a lane masking layer to recognize lane candidates, generate the lane mask, and generate a masked feature map; instructing a convolutional layer to generate a lane feature map; instructing a first FC layer to generate a softmax score map and a second FC layer to generate lane parameters; and backpropagating loss values outputted from a multinomial logistic loss layer and a line fitting loss layer, to thereby learn parameters of the FC layers, and the convolutional layer. Thus, lanes at distance can be detected more accurately.