Multitask Human Detection With Semantic Centers Under Occlusion
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional video surveillance and behavior detection systems rely on separate, task-specific detectors for human features, leading to inefficiencies, inconsistencies, and inaccuracies, especially in crowded scenes or under occlusion, due to reliance on geometric bounding box centers and lack of shared context.
Innovation Solution
A multitask detection model that integrates body detection, posture estimation, and semantic association into a single system, using semantic centers derived from shoulder and hip joint locations, and a two-stage regression architecture for improved anatomical accuracy.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Ease of operation
If separate task-specific detectors are used for human features, then each component can be executed independently, but inconsistencies and errors occur in grouping body parts into coherent subject representations
Solution Approach 1:
The patent combines multiple task-specific detectors (face detection, head detection, body detection, pose estimation) into a unified multitask detection framework. The system processes an input image through a single model that simultaneously outputs detections for multiple body parts and their associations, eliminating the need for separate independent executions while maintaining consistency through shared feature representations and joint optimization.
Solution Approach 2:
The unified detection model performs multiple functions simultaneously: detecting faces, heads, bodies, and posture points, while also determining their associations and visibility states. This multi-functional approach allows a single model to replace multiple specialized detectors, improving both efficiency and consistency in subject representation.
2Ease of manufacture
If bounding-box-based localization is used, then localization is simple to implement, but the geometric center fails to reflect a person's true anatomical position
Solution Approach 1:
The patent replaces the mechanical/geometric approach of using bounding box centers with a semantic/anatomical approach. Instead of computing the geometric center of a bounding box, the system uses a pose estimation model to detect anatomical landmarks (shoulders, hips, head) and computes the semantic center based on these biologically meaningful reference points, significantly improving anatomical position accuracy.
3Ease of manufacture
If conventional pose estimation models assume full-body visibility, then the models are simple to train, but they fail to produce useful results under occlusion or partial visibility
Solution Approach 1:
The patent modifies the output parameters of the pose estimation model to include visibility state predictions for each detected body part. Instead of only outputting location coordinates, the model simultaneously predicts whether each body part is visible or occluded, allowing it to handle partial visibility scenarios gracefully by indicating which parts are unreliable rather than producing incorrect predictions.
Solution Approach 2:
The system performs preliminary detection of body parts and their visibility states before final association and subject grouping. By first identifying which body parts are visible and reliable, the system can then use this information to guide the association process, excluding occluded or unreliable detections from final subject representations.
4Measurement precision
If multiple separate detectors are used for different body parts, then each detector can be optimized for its specific task, but the system becomes complex and computationally inefficient
Solution Approach 1:
The patent merges multiple separate detectors into a single unified multitask detection model. This model shares common feature extraction layers and simultaneously performs face detection, head detection, body detection, and pose estimation in one forward pass, reducing computational overhead while maintaining task-specific accuracy through specialized output heads for each detection type.
Data Source
AI summary
A system or method for multitask detection of human subjects in images. The system is configured to receive a captured image comprising one or more human subjects and to detect, using a pre-trained multitask detection model, a plurality of body portions for each subject, including at least a head, a body, and multiple posture body points. For each detected body portion, the system determines a semantic center representing an anatomically consistent location. The system further computes a plurality of vectors between the semantic centers of the body portions. Using these vectors, the system associates body portions belonging to the same human subject through part-to-part matching. For each subject, the system generates a bounding box annotation that encloses the associated body portions within the image.