Surveillance Event Rules From Operator Prompts and Context
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing surveillance systems face challenges with complexity, increased computational costs, and legal compliance burdens due to fine-tuning and reinforcement learning techniques, and rule-based solutions are time-consuming and require regular updates.
Innovation Solution
A method that utilizes a machine learning model, specifically a Large Language Model, to generate rules based on operator feedback and contextual knowledge, reducing the need for fine-tuning and reinforcement learning, and allowing for adaptive rule generation directly within the surveillance system.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If fine-tuning techniques are used to adapt models to specific tasks, then model performance is improved, but system complexity and computational costs increase
Solution Approach 1:
The patent extracts the model adaptation process from the surveillance system by using external generative AI models that can be independently trained and updated. The VMS calls these external models via API without requiring fine-tuning within the surveillance system itself, thus improving model performance while keeping system complexity manageable.
Solution Approach 2:
The patent introduces an intermediary layer (the generative AI model interface and prompt management system) between the surveillance data and the analysis engine. This intermediary handles the complexity of model adaptation externally, allowing the core VMS to remain simple while still achieving high-performance model adaptation through external services.
2Adaptability or versatility
If fine-tuning is performed within the installation, then model adaptation is achieved, but resource requirements and operational disruption increase
Solution Approach 1:
The computationally intensive model training and fine-tuning operations are extracted from the surveillance installation and performed externally. The VMS system calls pre-trained or externally fine-tuned models via API, eliminating the need for local computational resources to be dedicated to model adaptation while maintaining adaptability.
Solution Approach 2:
The patent performs model training and fine-tuning in advance (before deployment to the surveillance system). Models are pre-trained on general data and then fine-tuned externally on specific surveillance tasks beforehand, so that when deployed to the VMS, they are already optimized and require minimal additional computational resources during actual surveillance operations.
3Adaptability or versatility
If reinforcement learning is used to optimize system outcomes, then system adaptability is improved, but resource requirements and complexity increase
Solution Approach 1:
The patent implements feedback mechanisms where operator corrections and event outcomes are used to refine prompts and adjust model behavior. The system learns from feedback by adjusting prompt engineering strategies and model parameters externally, achieving adaptability without implementing complex reinforcement learning algorithms within the surveillance system itself.
4Reliability
If rules-based solutions are implemented, then event determination is improved, but operator time and maintenance burden increase
Solution Approach 1:
The patent enables the system to automatically generate and adjust its own rules through generative AI models. The models analyze surveillance data and automatically create event determination rules based on learned patterns, eliminating the need for operators to manually define and maintain complex rule sets while maintaining high reliability in event determination.
Solution Approach 2:
The patent dynamically adjusts rule parameters based on contextual information and operator feedback. Instead of fixed rules, the system modifies rule parameters (such as thresholds, time windows, and detection criteria) adaptively using generative AI, improving event determination reliability while reducing operator intervention time.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A computer-implemented method 40 for augmenting a surveillance system configured to receive a data stream from a detection device, analyse the data stream, generate content metadata about the content of the data stream based on the analysis of the data stream, and determine an event for display to a surveillance system operator based on the content metadata, the method comprising the steps of: receiving a surveillance system operator input 42 comprising a natural language text input; accessing contextual knowledge 43 of the surveillance system from a contextual knowledge source; determining an input 44 to a machine learning model based on the received surveillance system operator input and the contextual knowledge of the surveillance system from the contextual knowledge source; generating one or more rules 46 by the machine learning model based on the determined input; and applying the one or more generated rules 48 for modifying the event displayed to the operator.