Edge Loss Segmentation for Pedestrian Detection

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional image segmentation methods using CNNs often miss edge parts during encoding and decoding processes, leading to inefficient edge detection and reconstruction, which affects the accuracy of segmentation results.

Innovation Solution

The proposed method employs a learning device with convolutional and deconvolutional layers, mask layers, and edge loss layers to generate edge feature maps and reinforced decoded feature maps, adjusting parameters through backpropagation to improve edge detection accuracy by reinforcing edge parts.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional CNN encoding and decoding processes are used for image segmentation, then the segmentation process is simple and fast, but edge parts are missed and edge detection accuracy deteriorates

Engineering Contradiction:
Improveedge detection accuracyVSAvoidsegmentation process complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent divides the segmentation process into multiple parallel branches: a main decoding path and multiple edge decoding paths. Each edge decoding path processes edge information separately using dedicated edge detection layers and edge refinement layers, then fuses them with the main decoding result. This segmentation of processing paths enables specialized edge detection without compromising overall segmentation speed.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent introduces edge detection layers and edge refinement layers as intermediary components between the encoding and final segmentation output. These intermediary layers specifically process edge information through gradient calculation and refinement operations, acting as mediators that enhance edge detection accuracy before the final segmentation decision is made.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If edge parts are reconstructed during decoding, then edge detection capability improves, but energy consumption increases

Engineering Contradiction:
Improveedge detection capabilityVSAvoidenergy for edge reconstruction
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The patent performs edge detection and refinement operations at multiple intermediate stages during the decoding process, rather than attempting to reconstruct all edge information from the final encoded representation. By detecting and refining edges progressively at each decoding stage, the system reduces the computational energy required for complete edge reconstruction while maintaining high edge detection capability.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent applies different processing qualities to different regions of the image: the main decoding path processes the entire image at standard resolution, while edge decoding paths focus computational resources only on edge regions through gradient-based detection and refinement. This local quality approach concentrates energy expenditure on edge areas where it is most needed, reducing overall energy consumption compared to uniform high-quality processing.

Inventive Principle:
Principle #3Local quality

3Manufacturing precision

If multiple decoding paths are used to improve edge detection, then segmentation accuracy improves, but computational complexity increases

Engineering Contradiction:
Improvesegmentation accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Manufacturing precisionVSDevice complexity

Solution Approach 1:

The patent merges multiple decoding paths (main decoding and multiple edge decoding paths) into a unified segmentation output through element-wise addition and fusion operations. The main decoded feature map and edge-refined feature maps are combined in a way that integrates edge information with overall segmentation context, achieving high segmentation accuracy while managing computational complexity through efficient fusion rather than separate independent processing.

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The patent implements a hierarchical edge decoding approach where only certain critical edge regions are processed through full edge detection and refinement paths, while other regions use simplified processing. This partial action strategy applies complex edge processing only where necessary (excessive action in critical areas) rather than uniformly across the entire image, balancing segmentation accuracy with computational complexity management.

Inventive Principle:
Principle #16Partial or excessive action

Data Source

PatentEP3686795B1Learning method and learning device for improving segmentation performance to be used for detecting events including pedestrian event, vehicle event, falling event and fallen event using edge loss and test method and test device using the same
Publication Date: 2023.08.09 STRADVISION
  • EP3686795B1 patent drawingFigure 1
  • EP3686795B1 patent drawingFigure 2
  • EP3686795B1 patent drawingFigure 3

AI summary

A learning method for improving a segmentation performance to be used for detecting events including a pedestrian event, a vehicle event, a falling event, and a fallen event using a learning device is provided. The method includes steps of: the learning device (a) instructing k convolutional layers to generate k encoded feature maps; (b) instructing k-1 deconvolutional layers to sequentially generate k-1 decoded feature maps, wherein the learning device instructs h mask layers to refer to h original decoded feature maps outputted from h deconvolutional layers corresponding thereto and h edge feature maps generated by extracting edge parts from the h original decoded feature maps; and (c) instructing h edge loss layers to generate h edge losses by referring to the edge parts and their corresponding GTs. Further, the method allows a degree of detecting traffic sign, landmark, road marker, and the like to be increased.