Image Processing Apparatus for Occlusion-Resistant Action Recognition
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing action recognition techniques suffer from decreased accuracy due to occlusions, where parts of a person's body are obscured, leading to incorrect detection and recognition of actions, particularly when patterns resembling body parts are detected in occluded areas.
Innovation Solution
An image processing apparatus that estimates a person's posture and action by specifying organ positions, determining the state of occlusion, and correcting action recognition results based on predefined conditions to prevent incorrect outputs.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If image combination or data extension methods are used to handle occlusion, then posture estimation accuracy for occluded organs is improved, but false detection of patterns resembling organs in occluded areas increases
Solution Approach 1:
The system uses detection result feedback to iteratively refine organ position detection. When patterns resembling organs are detected in occluded areas, the feedback mechanism allows the system to re-evaluate and correct false detections by considering contextual information from multiple images and detection results, thereby maintaining high accuracy while reducing false positives
Solution Approach 2:
The system performs preliminary occlusion detection and marking before final organ detection. By identifying occluded areas in advance and applying appropriate processing or exclusion, the system prevents false detections from occurring in the first place, thus maintaining both accuracy and reliability
2Loss of information
If organ detection is performed in occluded areas, then complete posture information can be acquired, but incorrect organ detection occurs due to patterns resembling organs
Solution Approach 1:
The system segments the image into occluded and non-occluded areas, applying different processing strategies to each. In non-occluded areas, standard detection is performed with high accuracy. In occluded areas, the system either uses alternative detection methods or excludes these regions from final results, preventing false detections while preserving information from reliable areas
Solution Approach 2:
The system introduces an intermediary occlusion detection mechanism that identifies and flags suspicious detection results in occluded areas. This intermediary layer acts as a filter between raw detection and final results, allowing the system to maintain completeness by including detected information while ensuring accuracy through additional verification
3Productivity
If action recognition is performed using posture estimation results, then action classification can be achieved, but incorrect actions are output when posture estimation is degraded
Solution Approach 1:
The action recognition system incorporates feedback from posture estimation quality metrics. When posture estimation accuracy is degraded due to occlusion, the feedback mechanism triggers alternative recognition strategies or flags the result for review, maintaining efficient automated processing while preventing incorrect action classifications
Solution Approach 2:
The system performs partial action recognition based on available reliable posture information rather than requiring complete posture data. When some body parts are occluded, the system can still recognize actions based on visible parts, maintaining productivity while accepting that recognition confidence may be reduced and requiring additional verification
Data Source
AI summary
An image processing apparatus includes one or more memories storing instructions, and one or more processors that, upon execution of the stored instructions, are configured to specify a position of an organ of a person in an image, acquire posture estimation information about a posture of the person based on the specified position of the organ, estimate an action of the person based on the acquired posture estimation information, determine a state of the person based on the acquired posture estimation information, and change information indicating the estimated action of the person in a case where the estimated action of the person is a predetermined action and it is determined that the person is in a predetermined state.