CNN Lane Segmentation Loss Optimization

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

In image segmentation for autonomous vehicles, lanes with a small number of pixels are often misjudged as non-lanes due to uneven pixel distribution, leading to blurred segmentation results.

Innovation Solution

A learning method that calculates embedding loss and multinomial logistic loss to enhance segmentation scores, ensuring that lanes with fewer pixels are accurately distinguished by backpropagating Softmax and embedding losses, thereby increasing inter-lane differences and reducing intra-lane variances.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional image segmentation is used, then processing speed is maintained, but lanes with small number of pixels are blurred or misjudged as non-lanes

Engineering Contradiction:
Improvelane detection accuracyVSAvoidsegmentation reliability for small pixel clusters
Core Design Contradiction:
Measurement precisionVSReliability

Solution Approach 1:

The patent segments the image into multiple clusters based on lane detection, and further segments each cluster into multiple regions. This hierarchical segmentation allows the system to handle both large and small pixel clusters effectively by applying different processing strategies to different segments, thereby improving detection accuracy for lanes with few pixels while maintaining overall processing efficiency

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent applies different quality standards and processing methods to different regions within clusters. Specifically, it divides clusters into multiple regions and applies region-specific processing to ensure that small pixel clusters receive appropriate attention and processing, preventing them from being blurred or misjudged while maintaining high reliability for the entire segmentation process

Inventive Principle:
Principle #3Local quality

2Measurement precision

If conventional segmentation is applied to images with uneven pixel distribution, then processing complexity is reduced, but segmentation scores fail to distinguish different lanes clearly

Engineering Contradiction:
Improvesegmentation score discriminationVSAvoidloss calculation complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent segments clusters into multiple regions and calculates loss values for each region separately. This segmentation approach enables precise discrimination of segmentation scores for different lanes by handling each region's characteristics individually, while the modular loss calculation structure manages complexity through systematic regional processing

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces multiple loss functions (first loss function for between-cluster differences, second loss function for within-cluster region differences) that change the parameter space of optimization. This multi-parameter approach enhances segmentation score discrimination by optimizing multiple aspects simultaneously, while the structured loss calculation framework manages the increased complexity

Inventive Principle:
Principle #35Parameter changes

3Measurement precision

If standard loss backpropagation is used, then training efficiency is maintained, but inter-lane differences and intra-lane variances are not optimized

Engineering Contradiction:
Improveinter-lane difference and intra-lane variance optimizationVSAvoidloss function complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent employs multiple loss functions that operate on different parameter levels: the first loss function optimizes differences between clusters, while the second loss function optimizes differences between regions within clusters. This multi-level parameter optimization achieves superior inter-lane difference and intra-lane variance optimization, with the structured loss calculation framework managing complexity through systematic hierarchical processing

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3686842B1Learning method, learning device, testing method and testing device
Publication Date: 2022.07.13 STRADVISION
  • EP3686842B1 patent drawingFigure 1
  • EP3686842B1 patent drawingFigure 2
  • EP3686842B1 patent drawingFigure 3

AI summary

A learning method for segmenting an image having one or more lanes is provided to be used for supporting collaboration with HD maps required to satisfy level 4 of autonomous vehicles. The learning method includes steps of: a learning device instructing a CNN module (a) to apply convolution operations to the image, thereby generating a feature map, and apply deconvolution operations thereto, thereby generating segmentation scores of each of pixels on the image; (b) to apply Softmax operations to the segmentation scores, thereby generating Softmax scores; and (c) to (I) apply multinomial logistic loss operations and pixel embedding operations to the Softmax scores, thereby generating Softmax losses and embedding losses, where the embedding losses is used to increase inter-lane differences among averages of the segmentation scores and decrease intra-lane variances among the segmentation scores, in learning parameters of the CNN module, and (II) backpropagate the Softmax and the embedding losses.