Video Activity Detection Using Transaction Log Segmentation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Human review of video feeds for monitoring compliance with safety regulations and detecting activities of concern is time-consuming and inefficient, and existing automated systems struggle with accuracy due to clutter, poor lighting, and the need for supervised learning with labeled data, limiting their effectiveness in real-world environments.
Innovation Solution
A method that parses video streams into segments related to temporally distinct transactions or pseudo transaction markers, extracts visual image features, and uses one-SVM classification to categorize segments as inliers or outliers, issuing alerts for events of concern without requiring supervised learning or ground truths, thereby distinguishing true events from false alarms.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If human review and analysis of video feeds is used to monitor compliance and detect activities of concern, then accuracy in determining human activities can be maintained, but the process becomes time-consuming and inefficient
Solution Approach 1:
The patent segments video streams into discrete segments associated with specific transactions or events. By dividing the continuous video feed into manageable portions linked to particular activities (e.g., scanning transactions, customer interactions), the system can efficiently process and analyze only relevant portions rather than reviewing entire video feeds manually, thus reducing time loss while maintaining accuracy through focused analysis.
2Productivity
If automated analysis systems are implemented to reduce time consumption, then productivity increases, but accuracy decreases due to clutter, poor lighting, and lack of supervised learning
Solution Approach 1:
The patent introduces an intermediary layer between automated video analysis and final activity determination. This intermediary consists of transaction logs and event data that provide contextual information about what should be occurring in the video segments. By comparing visual features against this intermediary reference data, the system achieves both automated processing efficiency and improved accuracy, overcoming the limitations of pure supervised learning approaches.
Solution Approach 2:
The patent replaces traditional mechanical supervised learning approaches with an unsupervised learning methodology. Instead of requiring manually labeled training data and complex model training, the system uses unsupervised feature extraction and comparison against transaction logs, substituting the mechanical supervised learning process with a more efficient unsupervised approach that maintains or improves accuracy while increasing productivity.
3Measurement precision
If supervised learning with labeled data is used to improve detection accuracy, then measurement precision increases, but device complexity and data requirements increase
Solution Approach 1:
The patent enables the system to determine human activities without requiring external labeled training data or complex supervised learning infrastructure. The unsupervised learning approach allows the system to self-determine activity patterns by analyzing visual features in context with transaction logs, eliminating the need for manual data labeling and complex model training processes, thus reducing device complexity while maintaining detection accuracy.
Data Source
AI summary
Human behavior alerts are determined from a video stream through application of video analytics that parse a video stream into a plurality of segments, wherein each of the segments are either temporally related to at least one of a plurality of temporally distinct transactions in an event data log; or they are each associated with a pseudo transaction marker if not temporally related to at least one of the temporally distinct transactions and an image analysis indicates a temporal correlation with at least one of the distinct transactions is expected. Visual image features are extracted from the segments and one-SVM classification is performed on the extracted features to categorize segments into inliers or outliers relative to a threshold boundary. Event of concern alerts are issued with respect to the inlier segments associated with the associated pseudo transaction marker.