Pose-Aware Group Identification for Crowded Image Processing
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current image processing techniques struggle to accurately detect individual people within tightly packed groups, often misidentifying groups as a single individual or failing to identify any individual at all, especially when there is no definable separation between individuals.
Innovation Solution
A system and method utilizing a deep machine-learning algorithm to process limb and pose attributes, assigning limb and pose assignments to individuals within a group, and providing metadata to uniquely identify each person through bounding boxes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If existing image processing approaches are used to detect people in images, then processing speed is maintained, but detection accuracy deteriorates when individuals are tightly packed together
Solution Approach 1:
The patent segments the detection task into multiple specialized components: a people detector for initial identification, a pose estimator for limb positioning, and a grouping module for separating individuals. This segmentation allows each component to specialize in specific aspects of the detection process, improving overall accuracy for tightly packed individuals while maintaining manageable system complexity through modular architecture.
Solution Approach 2:
The patent introduces pose estimation as an additional dimensional layer of information beyond simple bounding boxes. By estimating limb positions and orientations in 2D space, the system creates a new dimension of data that enables differentiation between closely packed individuals, transforming the detection problem from basic object recognition to pose-aware identification.
2Measurement precision
If existing approaches detect groups of people, then processing simplicity is maintained, but individual identification capability is lost
Solution Approach 1:
The patent performs preliminary pose estimation and limb detection on the entire group before proceeding to individual separation and identification. By pre-processing the image to extract pose information for all detected people, the system prepares differentiated data structures that facilitate faster individual identification in subsequent steps, reducing overall processing time despite the enhanced accuracy requirements.
Solution Approach 2:
The patent introduces pose estimates and limb detections as intermediary data structures that bridge the gap between group-level detection and individual-level identification. These intermediaries contain detailed spatial information about each person's body configuration, enabling accurate individual separation without requiring excessive computational resources in the final identification stage.
3Measurement precision
If no separation is present between individuals in groups, then detection difficulty increases, but existing approaches still fail to identify any individuals
Solution Approach 1:
The patent applies local quality analysis by examining specific local features such as limb positions, body orientations, and relative distances between body parts of adjacent individuals. Instead of relying on global image characteristics, the system analyzes local pose qualities that remain distinctive even when individuals are tightly packed, enabling differentiation based on unique local body configurations.
Solution Approach 2:
The patent changes the detection parameters from basic bounding box coordinates to detailed pose estimation parameters including joint positions, limb orientations, and body angles. This parameter transformation provides additional degrees of freedom for distinguishing between individuals, as pose parameters capture subtle variations in body configuration that persist even when spatial separation is minimal.
Data Source
AI summary
Cameras capture time-stamped images of predefined areas. At least one image includes a representation of multiple individuals in a group of individuals. Attributes retained with each individual are combined with a limb and pose recognition to properly identify each individual of the group within the image.