Object-Aware Temporal Anomaly Detection for Edge Video Streams

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing anomaly detection systems in object recognition are computationally complex, fail to detect negative anomalies, and lack real-world knowledge, particularly in edge devices with limited computational resources, leading to inefficiencies and inaccurate detection of temporal patterns.

Innovation Solution

A system that incorporates real-world knowledge, such as time, day, and weather, to detect temporal pattern anomalies using a convolutional autoencoder and clustering modules, enabling efficient detection of both positive and negative anomalies on edge devices with limited resources.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If existing anomaly detection systems use convolutional layers and clustering to detect visual anomalies, then they can identify out-of-place objects, but they fail to detect negative anomalies and cannot recognize long-term temporal patterns

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoiddetection capability scope
Core Design Contradiction:
Measurement precisionVSAdaptability or versatility

Solution Approach 1:

The system performs preliminary object recognition and classification before anomaly detection, using pre-trained object recognition models to identify and label objects in video frames. This preliminary action enables the system to understand what objects should be present and when, forming the basis for detecting both positive and negative anomalies including temporal pattern violations

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system adds temporal dimension to anomaly detection by analyzing video sequences over time rather than individual frames. It incorporates time-of-day, day-of-week, and historical pattern information to detect anomalies that violate temporal expectations, such as objects appearing at wrong times or missing expected recurring events

Inventive Principle:
Principle #17Another dimension (Dimensionality change)

2Measurement precision

If deep CNN with many layers is used for fine-grained object classification, then classification accuracy is improved, but computational complexity and memory access requirements increase significantly

Engineering Contradiction:
Improveobject classification accuracyVSAvoidcomputational complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system extracts only the necessary computational components for the specific anomaly detection task rather than using full deep CNN architectures. It uses pre-trained object recognition models that have been extracted and adapted for edge device deployment, removing unnecessary layers and operations while retaining classification accuracy for relevant object categories

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system employs lightweight, optimized object recognition models designed for resource-constrained edge devices. These models use quantized weights, reduced precision arithmetic, and optimized inference engines to achieve acceptable accuracy with significantly reduced computational requirements compared to full-precision deep CNNs

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

3Measurement precision

If existing anomaly detection systems process video frames through convolutional layers, then they can produce feature representations, but they consume excessive computational resources and electricity on edge devices

Engineering Contradiction:
Improvefeature representation qualityVSAvoidenergy consumption
Core Design Contradiction:
Measurement precisionVSUse of energy by moving object

Solution Approach 1:

The system replaces traditional convolutional neural network processing with optimized inference approaches suitable for edge devices. It uses pre-computed feature representations from object recognition models and compares them directly against temporal patterns, avoiding repeated convolutional operations and reducing energy consumption while maintaining detection accuracy

Inventive Principle:
Principle #28Mechanics substitution (Replace mechanical system)

4Measurement precision

If existing anomaly detection systems cluster video frames to identify anomalies, then they can detect out-of-place objects, but they cannot provide text labels identifying the anomalous objects

Engineering Contradiction:
Improveanomaly identification accuracyVSAvoidobject identification information
Core Design Contradiction:
Measurement precisionVSLoss of information

Solution Approach 1:

The system incorporates feedback from object recognition labels into the anomaly detection process. By continuously monitoring identified objects and their temporal patterns, the system can provide meaningful text labels for anomalies such as 'person detected at wrong time' or 'expected object missing', enhancing interpretability and information retention

Inventive Principle:
Principle #23Feedback

Data Source

PatentEP3850502B1Method for detection of temporal pattern anomalies in video streams
Publication Date: 2025.10.01 CHIRAL SOFTWARE INC
  • EP3850502B1 patent drawingFigure 1
  • EP3850502B1 patent drawingFigure 2
  • EP3850502B1 patent drawingFigure 3A~3B

AI summary

A method for detecting temporal pattern anomalies in a video stream includes detecting an object in a current frame of the video stream, generating a processed current frame that contains the detected object, generating a feature representation of the processed current frame, clustering the feature representation in one or more primary clusters in a clustering space of the primary class, generating an information vector of the feature representation, that includes information regarding the primary class, the sub-class and one or more external factors associated with the feature representation, clustering each information vector into one or more secondary clusters, and reporting a next frame as an anomaly when a corresponding information vector is positioned outside a secondary cluster of a feature presentation of a previous frame.