Edge Visual Anomaly Detection via Contextual ML Filtering
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Efficiently analyzing vast volumes of visual data to detect anomaly events in scenes is challenging due to the high computational resources required, which can be costly and inefficient, especially when edge nodes have limited resources and remote servers are far away, leading to bandwidth and latency issues.
Innovation Solution
A system that uses an edge node with a limited resources classifier to identify potential anomaly events based on contextual attributes, applying a trained context-based Machine Learning model to compute anomaly scores, and transmitting only relevant images to a remote server for further analysis using high-performance tools.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If visual data is analyzed using high-performance tools on remote servers, then measurement precision and reliability of anomaly detection are improved, but loss of energy and bandwidth usage increase
Solution Approach 1:
The system segments the visual data processing workflow into two stages: edge nodes perform preliminary analysis using lightweight classifiers to identify potential anomalies, and only promising candidates are transmitted to remote servers for high-performance verification. This segmentation reduces bandwidth usage while maintaining detection accuracy.
Solution Approach 2:
Edge nodes perform preliminary filtering and classification of visual data before transmission to remote servers. By pre-processing data locally and transmitting only relevant anomalies, the system reduces energy consumption and bandwidth usage while preserving the ability to perform accurate analysis when needed.
2Measurement precision
If visual data is analyzed using high-performance tools on remote servers, then measurement precision and reliability of anomaly detection are improved, but loss of time due to data transmission increases
Solution Approach 1:
The system performs preliminary anomaly detection at edge nodes before transmission to remote servers. This preliminary action filters out non-anomalous data, so only relevant images requiring high-performance analysis are transmitted, significantly reducing transmission time while maintaining detection accuracy.
Solution Approach 2:
The system applies different processing qualities to different data: edge nodes use lightweight classifiers for initial screening, while remote servers apply high-performance tools only to promising candidates. This local quality differentiation reduces transmission time while preserving measurement precision where needed.
3Loss of energy
If edge nodes use limited resources classifiers, then loss of energy and device complexity are reduced, but measurement precision of anomaly detection deteriorates
Solution Approach 1:
The system segments the classification task across two levels: edge nodes use limited-resources classifiers for initial screening to identify potential anomalies, and remote servers use high-performance tools for final verification. This segmentation allows energy-efficient preprocessing while maintaining overall detection precision.
Solution Approach 2:
The edge node acts as an intermediary between imaging sensors and remote servers, performing preliminary classification with limited resources and forwarding only promising candidates to the server. This intermediary role reduces computational cost at the edge while preserving measurement precision through server-side verification.
Data Source
AI summary
Disclosed herein are methods and systems for visually identifying anomaly events, comprising an edge node configured for applying a limited resources classifier to a plurality of images captured by imaging sensor(s) deployed to monitor a certain scene relating to a certain area to classify object(s) detected in the images, applying a trained context based Machine Learning (ML) model to classification data generated by the limited resources classifier to compute an anomaly score for potential anomaly event(s) relating to the detected object(s) based on one or more contextual attributes associated with the certain scene and transmitting one or more of the images to a remote server in case the anomaly score exceeds a threshold. The remote server is configured to further apply high performance visual analysis tool(s) to visually analyze the received image(s) in order to identify the one or more potential anomaly events.