Dynamic Video Stream Ranking for Anomaly Detection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current surveillance systems face challenges in monitoring numerous cameras and sensors with limited operators, as they are preconfigured to detect specific events, missing occurrences not predefined, and struggle to prioritize and predict relevant events in real-time.
Innovation Solution
A system that dynamically orders video streams using a neural network to extract features, calculate abnormality scores, and prioritize events based on current and past scores, allowing for real-time updates and user feedback to improve detection and classification of anomalies.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If preconfigured event types are used for detection, then specific event detection capability is improved, but detection of unconfigured events deteriorates
Solution Approach 1:
The system employs a universal anomaly detection model that can identify multiple types of events without requiring separate preconfigured detectors for each event type. The neural network processes video frames and extracts features that represent various abnormal behaviors, enabling the system to detect intrusion, motion, people count, crowd management, camera tamper, and other events through a single multi-functional detection mechanism.
Solution Approach 2:
The system changes the detection parameter from specific event type classification to general anomaly scoring. Instead of configuring detectors for specific events, the system calculates abnormality scores based on extracted video features and presents events ranked by their anomaly level, allowing detection of both preconfigured and unconfigured event types through parameter transformation.
2Reliability
If all video streams are displayed to operators, then complete monitoring coverage is improved, but operator attention and efficiency deteriorate
Solution Approach 1:
The system applies local quality by differentiating the presentation of video streams based on their anomaly characteristics. Instead of uniform display, streams with higher abnormality scores are prioritized and presented more prominently to operators, while normal streams are deprioritized. This creates non-uniform display quality that matches the operational need for focused attention on critical events.
Solution Approach 2:
The system performs preliminary action by pre-calculating and ranking video streams according to their abnormality scores before presentation to operators. The ranking is computed in advance based on extracted features and abnormality calculations, so that when operators view the streams, they are already ordered by relevance, eliminating the need for operators to manually prioritize streams during monitoring.
3Reliability
If video analytics engines detect only preconfigured events, then detection reliability for known events is improved, but detection of novel anomalies deteriorates
Solution Approach 1:
The system employs self-service by using unsupervised anomaly detection that does not require manual configuration of event types or thresholds. The neural network automatically learns normal patterns from video data and independently identifies deviations as anomalies, enabling the system to detect both known and novel events without human intervention for configuration or reconfiguration.
Solution Approach 2:
The system implements dynamics by making the detection model adaptable and evolving. The abnormality detection mechanism can identify new types of anomalies as they appear in the video stream, and the system dynamically adjusts to detect previously unseen event types. This dynamic capability allows the system to maintain reliability for known events while simultaneously gaining versatility for novel anomalies.
Data Source
Figure 1
Figure 2
Figure 3
AI summary
A system and method for dynamically ordering video streams on display in a command and control system according to video content, including obtaining a plurality of video streams captured by a plurality of respective video cameras; extracting features from the plurality of video streams; calculating abnormality scores for the extracted features, wherein each abnormality score measures a likelihood that a corresponding feature is an abnormal feature; calculating a current rank for ordering each of the plurality of video streams based on current and past abnormality scores of the features extracted from the video stream; and displaying a subgroup of the plurality of video streams having the highest subset of current ranks based on abnormality scores.