Abnormal Action Detector Using Pixel-by-Pixel Auto-Correlation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional abnormal action recognition systems fail to accurately identify the position of an object performing an abnormal action due to position invariance of cubic higher-order local auto-correlation features and suffer from lower detection accuracy with multiple objects or types of objects.
Innovation Solution
The system generates inter-frame differential data, extracts feature data using cubic higher-order local auto-correlation, calculates an index of abnormality based on a principal component subspace, and classifies pixel positions using a clustering approach to determine abnormal actions on a pixel-by-pixel basis, allowing for accurate detection of abnormal actions regardless of object position and type.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Productivity
If cubic higher-order local auto-correlation features are extracted from the entire screen, then detection speed is improved and labor is reduced, but the position of the object performing abnormal action cannot be identified
Solution Approach 1:
The screen is divided into multiple regions, and feature extraction is performed for each region separately. This segmentation allows the system to maintain position information while still using automated feature extraction, resolving the contradiction between detection speed and position identification capability.
Solution Approach 2:
The patent applies different processing approaches to different regions of the screen. By performing pixel-by-pixel feature extraction in specific regions where abnormal actions are suspected, the system preserves local position information while maintaining overall detection efficiency.
2Extent of automation
If cubic higher-order local auto-correlation features are used for abnormal action detection, then detection automation is improved, but detection accuracy decreases when there are multiple objects or types of objects
Solution Approach 1:
The system dynamically adjusts the detection approach based on the number and types of objects detected. When multiple objects are present, the system switches to object-specific feature extraction and tracking, maintaining high detection accuracy while preserving automation. This dynamic adaptation resolves the contradiction between automation extent and detection precision.
Solution Approach 2:
The patent changes detection parameters such as feature extraction granularity and tracking intensity based on scene complexity. For simple scenes with single objects, coarse feature extraction is used for high speed and automation. For complex scenes with multiple object types, finer-grained pixel-by-pixel extraction is applied to maintain accuracy, thus resolving the contradiction.
3Measurement precision
If pixel-by-pixel feature extraction is performed, then position identification accuracy is improved, but processing complexity and time increase
Solution Approach 1:
The system performs preliminary detection at a coarser level first to identify regions of interest where abnormal actions occur. Only in these specific regions is pixel-by-pixel feature extraction performed, significantly reducing overall processing complexity while maintaining high position identification accuracy where it matters most.
Solution Approach 2:
Instead of performing pixel-by-pixel extraction across the entire screen, the patent applies this computationally intensive method only to partial regions where abnormal actions are detected or suspected. This partial application maintains position identification accuracy while avoiding the excessive processing complexity that would result from full-screen analysis.
Data Source
AI summary
An abnormal action detector is provided for accurately detecting the position of an object together with an abnormal action using a cubic higher-order local auto-correlation feature. The abnormal action detector comprises a computer which generates inter-frame differential data from moving image data, extracts cubic higher-order local auto-correlation feature data on a pixel-by-pixel basis, adds the feature data for pixels within a predetermined range including each of pixels, calculates an index indicative of abnormality of the added feature data with respect to a subspace indicative of a normal action, determines an abnormality when the index is larger than a predetermined value, and outputs the position of a pixel at which the abnormality is determined. The computer further finds a subspace which exhibits a normal action from past feature data in accordance with a principal component analysis approach. The abnormal action detector is capable of determining abnormality on a pixel-by-pixel basis and correctly detecting the position of an object which has shown an abnormal action.