Real-Time Object Detection via Continual Learning Feature Fingerprinting

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Autonomous vehicles face challenges in detecting new objects that were not present during their training, leading to inefficient object detection, as existing CNN-based systems require costly updates and parameter transfers.

Innovation Solution

An object detecting system that updates in real-time by generating feature maps and query vectors from input images, comparing them to feature fingerprint databases, and using short-term and long-term update units to refine and expand the database entries, allowing for the detection of untrained classes.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Adaptability or versatility

If CNN parameters are updated regularly to detect new objects, then object detection capability is improved, but learning cost and parameter transfer cost increase

Engineering Contradiction:
Improveobject detection capabilityVSAvoidlearning cost
Core Design Contradiction:
Adaptability or versatilityVSLoss of energy

Solution Approach 1:

The system enables autonomous vehicles to perform self-learning by continuously generating feature maps and query vectors from captured images, comparing them against stored feature fingerprints, and automatically updating their detection database without external intervention or costly retraining processes

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The patent implements continuous learning where the object detection system operates indefinitely, constantly updating its knowledge base with new object types encountered during operation, rather than requiring periodic complete retraining cycles

Inventive Principle:
Principle #20Continuity of useful action

2Adaptability or versatility

If CNN parameters are updated regularly to detect new objects, then object detection capability is improved, but parameter transfer cost increases

Engineering Contradiction:
Improveobject detection capabilityVSAvoidparameter transfer cost
Core Design Contradiction:
Adaptability or versatilityVSQuantity of substance

Solution Approach 1:

The system extracts only the essential feature representations (query vectors) from full images and stores compact feature fingerprints in the database, transferring only these condensed representations rather than complete model parameters, significantly reducing data transfer requirements

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

Instead of transferring complete CNN parameters, the system creates and transfers simplified copies of object features in the form of query vectors and feature fingerprints, which capture essential object characteristics without the complexity of full model parameters

Inventive Principle:
Principle #26Copying

3Reliability

If CNN is trained on fixed training data sets, then initial object detection performance is achieved, but inability to detect new objects arises over time

Engineering Contradiction:
Improveinitial object detection performanceVSAvoiddetection of new objects
Core Design Contradiction:
ReliabilityVSAdaptability or versatility

Solution Approach 1:

The system transitions from a static, fixed training model to a dynamic, evolving detection system that continuously adapts to new object types by learning from images captured during operation, updating its feature database in real-time

Inventive Principle:
Principle #15Dynamics

Solution Approach 2:

The system uses feedback from captured images and their corresponding feature maps to continuously refine and expand its detection capabilities, comparing new object features against existing fingerprints and updating the database based on observed patterns

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3690710A1Method for providing object detecting system capable of updating types of detectable classes in real-time by using continual learning and device using the same
Publication Date: 2020.08.05 STRADVISION
  • EP3690710A1 patent drawingFigure 1
  • EP3690710A1 patent drawingFigure 2
  • EP3690710A1 patent drawingFigure 3

AI summary

A method for updating an object detecting system to detect objects with untrained classes in real-time is provided. The method includes steps of: (a) the object detecting system, if at least one input image is acquired, instructing a recognizer included therein to generate a specific feature map, and to generate a specific query vector; (b) the object detecting system instructing a similarity determining unit (i) to compare the specific query vector to data vectors, to thereby calculate each of first similarity scores between the specific query vector and each of the data vectors, and (ii) to add a specific partial image to an unknown image DB, if a specific first similarity score is smaller than a first threshold value; (c) the object detecting system, if specific class information is acquired, instructing a short-term update unit to generate a specific short-term update vector, and update the feature fingerprint DB.