Video Action Detection Using Reversibility and Speed Constraints
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing computer vision techniques struggle to accurately differentiate between reversible and temporally dependent actions, such as walking and running, or actions like parking and driving a car, which are often treated as the same, affecting the AI model's performance in tasks.
Innovation Solution
A method and system that apply logical constraints through action prediction labels, bounding boxes, classification loss, localization loss, reversibility action loss, and speed action loss to differentiate between similar but reversible actions, using transformers and neural networks to learn motion patterns and enforce boundaries between classes.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional AI models are used for action detection, then the model structure is simple, but the model cannot accurately differentiate between reversible actions (e.g., walking vs. running, parking vs. driving)
Solution Approach 1:
The patent segments the action detection problem into multiple independent loss functions: classification loss for basic action recognition, localization loss for spatial positioning, reversibility loss for distinguishing reversible actions, and speed loss for tempo-dependent actions. This segmentation allows each loss function to specialize in one aspect, improving overall accuracy without requiring a fundamentally more complex model architecture.
Solution Approach 2:
The patent introduces new parameters (reversibility labels, speed labels) to the existing action detection framework. By augmenting the label space with these additional parameters, the model can differentiate between reversible actions and tempo-dependent actions without changing the core network structure, thus improving measurement precision while maintaining reasonable complexity.
2Reliability
If the model treats reversible actions as the same, then the classification process is simpler, but the model fails to capture temporal dependencies and action distinctions
Solution Approach 1:
The patent separates the classification reliability problem into distinct loss functions: classification loss for basic accuracy and reversibility loss for distinguishing reversible actions. This segmentation ensures that each loss function contributes to a specific aspect of reliability without creating unnecessary complexity in the overall system.
Solution Approach 2:
The reversibility loss function provides feedback to the model during training by comparing predicted actions with ground truth reversibility labels. This feedback mechanism enables the model to learn temporal dependencies and action distinctions iteratively, improving classification reliability through continuous refinement without requiring overly complex architecture changes.
3Measurement precision
If the model ignores action speed, then the processing is faster, but the model cannot distinguish between tempo-dependent actions like walking and running
Solution Approach 1:
The patent adds speed labels as an additional parameter to the action detection framework. By incorporating this new parameter into the loss function (speed loss), the model can distinguish between tempo-dependent actions like walking and running. The parameter change approach allows the model to capture temporal information without fundamentally altering the processing pipeline, maintaining a balance between precision and complexity.
Data Source
AI summary
Systems and methods for action detection are provided. The systems and methods include generating action prediction labels and bounding boxes for objects detected in video frames and comparing the action prediction labels and the bounding boxes with corresponding ground labels, determining a classification loss and a localization loss, and determining a reversibility action loss by comparing the action prediction labels with known actions indices and logical constraints and a speed action loss by comparing the action prediction labels with known actions indices and logical constraints. The systems and methods further include combining the classification loss, localization loss, reversibility action loss, and speed action loss to evaluate a total loss and selecting the action prediction with a lowest total loss as an action assertion and performing reactionary actions in a connected device in response to the action assertion.