Video Anomaly Detection Using Human-Scene Semantic Drift

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Existing video anomaly detection methods struggle with false alarms due to focusing on global features, missing significant scene and human features, and failing to accurately detect subtle anomalies.

Innovation Solution

A method and system that separately extracts human-related and scene-related features, determines semantic drift between them, and combines these features into an augmented feature map using an encoder-decoder model and attention mechanisms to enhance anomaly detection.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If global snippet/frame-level features are used for anomaly detection, then the detection coverage is improved, but false alarms increase and significant features are missed

Engineering Contradiction:
Improveanomaly detection accuracyVSAvoidloss of significant human-scene features
Core Design Contradiction:
ReliabilityVSLoss of information

Solution Approach 1:

The video is segmented into human-related features and scene-related features, which are then processed separately through dedicated encoder-decoder models. This segmentation allows each feature type to be analyzed independently, preserving significant features that would otherwise be lost in global feature aggregation.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

Semantic drift is introduced as an intermediary mechanism that bridges human-related features and scene-related features. By computing semantic drift between these separately extracted features, the system identifies anomalies caused by inconsistencies between human actions and scene context, thereby reducing false alarms while maintaining detection accuracy.

Inventive Principle:
Principle #24Intermediary (Mediator)

2Ease of manufacture

If single video-level binary labels are used for training, then the training process is simplified, but frame-level temporal boundaries for abnormal events cannot be accurately determined

Engineering Contradiction:
Improvetraining process simplicityVSAvoidframe-level temporal boundary precision
Core Design Contradiction:
Ease of manufactureVSMeasurement precision

Solution Approach 1:

The system performs preliminary extraction of human-related features and scene-related features separately before combining them. This preliminary action enables the subsequent computation of semantic drift at frame-level, which accurately determines temporal boundaries of abnormal events while maintaining the simplicity of video-level binary label training.

Inventive Principle:
Principle #10Preliminary action

3Loss of information

If only global features are focused on, then the overall video understanding is improved, but subtle local anomalies are missed and irrelevant features dominate

Engineering Contradiction:
Improveoverall video understandingVSAvoiddetection of subtle local anomalies
Core Design Contradiction:
Loss of informationVSMeasurement precision

Solution Approach 1:

The patent applies local quality by treating human-related features and scene-related features differently through separate encoder-decoder models. Each model is optimized for its specific feature type, enabling precise detection of subtle local anomalies in human actions while maintaining overall video understanding through the combination of both feature streams.

Inventive Principle:
Principle #3Local quality

Data Source

PatentEP4664419A1Video anomaly detection method and system
Publication Date: 2025.12.17 TOYOTA JIDOSHA KK
  • EP4664419A1 patent drawingFigure 1
  • EP4664419A1 patent drawingFigure 2
  • EP4664419A1 patent drawingFigure 3

AI summary

A computer-implemented method for detecting an anomaly in a video, comprising: - obtaining human-related features (FH) and scene-related features (FS) extracted from the video; - determining a semantic drift (FD) between the human-related features (FH) and the scene-related features (FS); - selectively combining the human-related features (FH) and the scene-related features (FS), based on the semantic drift (FD), into an augmented feature map (F**); - detecting the anomaly based on the augmented feature map (F**).