Handed Activity Recognition Using Mirror-Flipped Training Data
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing deep learning-based activity recognition systems face challenges in distinguishing between left- and right-handed activities, which doubles the number of classes and increases complexity, and are sensitive to imbalanced population distributions, especially when using public datasets with a majority of right-handed individuals.
Innovation Solution
An activity recognition system that leverages left-right mirror symmetry by flipping input data horizontally and using a single handed activity classifier to infer both left- and right-handed activities, with a training method that balances the dataset by flipping all training instances to simulate left-handed activities.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If the system distinguishes between left- and right-handed activities by doubling the number of classes, then the recognition accuracy for asymmetric activities is improved, but the device complexity and training cost increase
Solution Approach 1:
The patent applies asymmetry by explicitly modeling left-handed and right-handed activities as separate classes with distinct features. The system captures asymmetric movements (e.g., left-footed vs right-footed kicking) through separate class labels and trained classifiers, enabling precise differentiation between-handed activities while maintaining manageable complexity through targeted feature extraction rather than treating all activities as symmetric
Solution Approach 2:
The system segments the activity recognition task by dividing activities into separate classes based on handedness. Instead of treating all activities uniformly, the patent segments the feature space and class space to handle asymmetric activities specifically, allowing the system to focus computational resources on distinguishing left- versus right-handed variants where needed
2Ease of manufacture
If the training dataset is sampled from public data with a majority of right-handed individuals, then the data acquisition process is simplified, but the classifier becomes sensitive to imbalanced population distribution
Solution Approach 1:
The patent applies parameter changes by transforming the class distribution parameters in the training data. The system modifies the label distribution from the natural imbalanced state (majority right-handed) to a balanced state where left- and right-handed activities are equally represented. This is achieved by flipping the handedness labels of a subset of training examples, thereby changing the statistical parameters of the training set to achieve equal representation and improve classifier reliability
3Measurement precision
If the system treats left- and right-handed activities as separate classes, then the recognition precision for asymmetric activities is improved, but the training database size and training procedure complexity increase
Solution Approach 1:
The system uses copying by creating synthetic training examples through horizontal flipping of existing video clips. Instead of requiring separate recordings of left-handed and right-handed activities, the patent takes right-handed activity examples, flips them horizontally, and uses these flipped copies as training data for left-handed activity classification. This copying approach doubles the effective training data for both-handed activities without requiring additional data collection
Data Source
AI summary
This disclosure describes an activity recognition system for asymmetric (e.g., left- and right-handed) activities that leverages the symmetry intrinsic to most human and animal bodies. Specifically, described is 1) a human activity recognition system that only recognizes handed activities but is inferenced twice, once with input flipped, to identify both left- and right-handed activities and 2) a training method for learning-based implementations of the aforementioned system that flips all training instances (and associated labels) to appear left-handed and in doing so, balances the training dataset between left- and right-handed activities.