Adaptive ML Model Training via Incident Timeline and Camera Retrieval

Resolve Bottlenecks,
Find Innovative SolutionsGenerate Solutions

Solution Overview

Problem

As the number of audio and video streams increases, it becomes difficult and time-consuming to train and update machine learning models to detect new situations, camera types, lighting conditions, and other parameters, leading to a decrease in the effectiveness of audio and video monitoring for public safety incidents.

Innovation Solution

A system and method for adaptive training of machine learning models using detected contextual public safety incident timeline entries, which involves correlating in-field incident timeline information with imaging camera databases to retrieve and use audio and video streams for training, allowing for continuous improvement of event detection models.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If the number of audio and video streams increases to improve coverage and detection capability, then the ability to identify situations of concern improves, but the time and complexity required to train and update machine learning models increases substantially

Engineering Contradiction:

Improveability to identify situations of concernVSAvoidtime to train and update models

Core Design Contradiction:

ReliabilityVSLoss of time

Solution Approach 1:

The system pre-organizes video streams into groups based on geographic location and incident type before training is needed. When an incident is detected, the system has already prepared relevant video groups, eliminating the need to search and organize streams during urgent model training situations.

Inventive Principle:

Principle #10Preliminary action

Solution Approach 2:

The system divides the large set of video streams into smaller, manageable groups based on geographic location and incident type. This segmentation allows the system to work with only relevant video streams during model training, reducing the computational burden and time required.

Inventive Principle:

Principle #1Segmentation

2Adaptability or versatility

If more machine learning models are deployed to detect various events and situations, then the comprehensiveness of incident detection improves, but the difficulty and time required to train and verify model outputs increases

Engineering Contradiction:

Improvecomprehensiveness of incident detectionVSAvoiddifficulty to train and verify models

Core Design Contradiction:

Adaptability or versatilityVSDevice complexity

Solution Approach 1:

The system creates a universal training framework that works across multiple incident types and camera locations. By organizing video streams into reusable groups and using standardized training procedures, the same system can train multiple different models without requiring separate complex processes for each.

Inventive Principle:

Principle #6Universality (Multi-functionality)

Solution Approach 2:

The system pre-categorizes video streams by incident type and location, so when new models need to be trained for specific incident types, the relevant training data is already organized and ready. This eliminates the need to manually search and organize video streams for each new model training task.

Inventive Principle:

Principle #10Preliminary action

3Measurement precision

If video streams from multiple camera locations are used to improve training data quality, then the accuracy of event detection improves, but the complexity of managing and retrieving relevant streams increases

Engineering Contradiction:

Improveaccuracy of event detectionVSAvoidcomplexity of managing video streams

Core Design Contradiction:

Measurement precisionVSDevice complexity

Solution Approach 1:

The system segments video streams by geographic location and incident type, creating organized groups that are easy to manage. This segmentation allows the system to efficiently retrieve only the relevant video streams needed for training specific models, rather than managing all streams as a single large set.

Inventive Principle:

Principle #1Segmentation

Solution Approach 2:

The system introduces an intermediate organization layer between the raw video streams and the training process. Video streams are pre-grouped by location and incident type, serving as an intermediary structure that simplifies the retrieval process and reduces the complexity of managing diverse video sources.

Inventive Principle:

Principle #24Intermediary (Mediator)

Data Source

PatentUS11417128B2Method, device, and system for adaptive training of machine learning models via detected in-field contextual incident timeline entry and associated located and retrieved digital audio and/or video imaging

Publication Date: 2022.08.16 MOTOROLA SOLUTIONS INC

AI summary

Receive first context information (FCI) including entered in-field incident timeline information values from an in-field incident timeline application and a time associated with an entry of the FCI values. Access a mapping that maps in-field incident timeline information values to events having a pre-determined threshold confidence of occurring and identify an event associated with the received FCI. Determine a location associated with the entry of the FCI and a time period associated with the entry of the FCI. Access a camera location database and identify cameras that have a field of view including the location during the time period. Retrieve audio and/or video streams captured by the cameras during the time period. And provide the audio and/or video streams to machine learning training modules corresponding to machine learning models for detecting the event in and/or video streams for further training of the machine learning models.