Image Recognition Target Matching Across Conflicting Class Estimates

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

Existing image recognition systems incorrectly estimate the class of recognition targets like motorcycles as pedestrians due to varying imaging angles and distances, leading to misidentification and inaccurate tracking.

Innovation Solution

An image recognition device with a determination unit that determines whether recognition targets of different classes are the same by checking if they are within a predetermined distance and using a majority rule or confidence levels to unify classes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If image recognition systems estimate class based on imaging angle and distance, then detection coverage is improved, but measurement precision deteriorates causing misidentification

Engineering Contradiction:

Improvedetection coverageVSAvoidclass estimation accuracy

Core Design Contradiction:

Adaptability or versatilityVSMeasurement precision

Solution Approach 1:

The system uses tracking information from previous frames as feedback to verify and correct class estimates. When a target is detected in the current frame, the system checks whether it matches the position and characteristics of tracked targets from previous frames, using this feedback to resolve classification ambiguities caused by varying imaging conditions.

Inventive Principle:

Principle #23Feedback

Solution Approach 2:

The system performs tracking preparation in advance by maintaining position information of detected targets across frames. This preliminary tracking data is ready before class estimation is needed, allowing the system to quickly resolve classification uncertainties by comparing current detections with pre-established tracking information.

Inventive Principle:

Principle #10Preliminary action

2Stability of the object's composition

If recognition targets are tracked across frames using position matching, then tracking continuity is improved, but reliability deteriorates when targets of different classes occupy similar positions

Engineering Contradiction:

Improvetracking continuityVSAvoidtarget identification accuracy

Core Design Contradiction:

Stability of the object's compositionVSReliability

Solution Approach 1:

The system adds temporal dimension to the tracking process by comparing not only spatial position but also temporal consistency of class estimates across multiple frames. By examining whether class estimates remain consistent over time, the system can distinguish between genuine targets and misidentified objects even when they occupy similar positions.

Inventive Principle:

Principle #17Another dimension (Dimensionality change)

Solution Approach 2:

The system introduces tracking ID as an intermediary element that mediates between position matching and class estimation. The tracking ID serves as a bridge that connects position information across frames while allowing class verification, enabling the system to maintain tracking continuity while ensuring identification reliability through cross-validation.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS20250292543A1Image recognition device, storage medium, and image recognition method

Publication Date: 2025.09.18 DENSO CORP

AI summary

An image recognition device acquires camera images captured by at least one camera, detects recognition targets in the camera images and estimates a class of the detected recognition targets respectively, and determines whether a plurality of recognition targets that are estimated to be of different classes are the same recognition targets.