ROI Detection Model Retraining Using Failure-Selected Video Frames

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional object detection systems face challenges in achieving high accuracy due to inadequate training data, leading to false positives and negatives, and are unable to effectively address these issues.

Innovation Solution

A method for object detection that involves selecting and re-training the ROI detection model using a tailored training dataset, which includes frames where the model fails to detect objects (false negatives) or incorrectly detects boundaries (false positives), utilizing ROI tracking and motion modeling to identify and correct these errors.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Quantity of substance

If all generated training data from on-premise raw videos is used, then the quantity of training data increases, but the detection model produces more false positives and false negatives

Engineering Contradiction:
Improvequantity of training dataVSAvoiddetection accuracy
Core Design Contradiction:
Quantity of substanceVSReliability

Solution Approach 1:

The patent extracts and removes redundant data from the training dataset. The system identifies and eliminates duplicate or highly similar video frames that do not provide additional learning value, keeping only diverse and informative samples. This extraction process reduces dataset size while maintaining or improving detection accuracy by removing harmful redundant information.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The patent applies local quality by selecting specific high-value frames for training rather than using all frames uniformly. The system evaluates each frame's information content and selects only those frames that provide unique or valuable learning signals, creating a non-uniform distribution of training samples where quality varies by frame importance rather than treating all data equally.

Inventive Principle:
Principle #3Local quality

2Reliability

If the training dataset is diversified with frames highlighting model failures, then detection accuracy improves, but the complexity of data selection and processing increases

Engineering Contradiction:
Improvedetection accuracyVSAvoiddata selection complexity
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent implements feedback by using the detection model's own performance to guide data selection. The system runs the model on validation data, identifies false positives and false negatives, and uses these error cases to select additional training frames. This closed-loop feedback mechanism automatically focuses training on weak areas without manual intervention, improving accuracy while managing complexity through automation.

Inventive Principle:
Principle #23Feedback

Solution Approach 2:

The system performs self-service by automatically identifying its own weaknesses and selecting training data to address them. The detection model evaluates itself on validation data, identifies problematic cases, and triggers selective data collection and retraining without external guidance. This self-directed improvement process reduces the need for complex external data curation while持续提升ding performance.

Inventive Principle:
Principle #25Self-service

Data Source

PatentEP4399689B1Object detection systems and methods including an object detection model using a tailored training dataset
Publication Date: 2026.02.11 TYCO FIRE & SECURITY GMBH
  • EP4399689B1 patent drawingFigure 1
  • EP4399689B1 patent drawingFigure 2
  • EP4399689B1 patent drawingFigure 3

AI summary

Disclosed herein is an object detection system, including apparatuses and methods for object detection. An implementation may include receiving a first image frame from an ROI detection model that generated a first ROI boundary around a first object detected in the first image frame and subsequently receiving a second image frame. The implementation further includes predicting, using an ROI tracking model, that the first ROI boundary will be present in the second image frame and then detecting whether the first ROI boundary is in fact present in the second image frame. The implementation includes determining that the second image frame should be added to a training dataset for the ROI detection model when detecting that the ROI detection model did not generate the first ROI boundary in the second image frame as predicted and re-training the ROI detection model using the training dataset.