Long-Distance Gesture Recognition With Adaptive Patch Selection
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing gesture recognition systems struggle to accurately recognize gestures from a distance greater than one meter due to small spatial regions of view and limited adaptability to variations in gestures performed by different users.
Innovation Solution
A machine learning system with a neural network architecture comprising a feature extractor, patch selector, and gesture recognizer, utilizing cross entropy losses for training, enables accurate gesture recognition from a distance by focusing on relevant patches and classifying gestures effectively.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If gesture recognition systems use standard sensor fields-of-view, then they can capture the entire scene, but they fail to recognize gestures accurately when subjects are more than one meter away because gestures occupy only small spatial regions
Solution Approach 1:
The patent divides the feature data into multiple patches and uses a patch selector to identify and focus on the specific patch containing the gesture. This segmentation approach allows the system to concentrate computational resources on the relevant gesture region rather than processing the entire large FOV, thereby improving recognition accuracy for distant subjects.
Solution Approach 2:
The patent extracts the selected patch containing the gesture from the larger feature data. By taking out only the relevant gesture portion and processing it separately through the gesture recognizer, the system achieves better recognition accuracy for small distant gestures without being overwhelmed by the entire scene.
2Adaptability or versatility
If gesture recognition systems are trained on fixed gesture sets, then they can recognize those specific gestures, but they cannot adapt to variations of gestures performed by different users
Solution Approach 1:
The patent implements a dynamic system where the patch selector and gesture recognizer can adapt to different users and gesture variations. The loss function combines cross-entropy losses from both the patch selector and gesture recognizer, enabling the system to learn and adapt to new gesture patterns while maintaining recognition of previously learned gestures, thus achieving both adaptability and reliability.
3Productivity
If gesture recognition systems process entire sensor data, then they have complete information, but they incur high computational load that reduces processing efficiency
Solution Approach 1:
The patent extracts only the selected patch containing the gesture from the full sensor data for detailed processing by the gesture recognizer. This extraction approach significantly reduces the amount of data requiring intensive computational processing, thereby improving processing efficiency and reducing computational complexity while maintaining recognition accuracy.
Solution Approach 2:
The patent segments the processing pipeline into a patch selection stage and a gesture recognition stage. By dividing the computational task this way, the system processes only relevant portions of the data at high computational cost, improving overall productivity while managing device complexity.
Data Source
AI summary
A computer-implemented system and method relate to gesture recognition. A machine learning system is trained using a training dataset of sensor data that include a set of gestures. The training dataset includes at least a first subset that displays a first gesture. Loss data is generated based on a first loss function that includes a first cross entropy loss and a second cross entropy loss. Parameters of the machine learning system are updated based on the loss data. The machine learning system is outputted and configured for gesture recognition of the set of gestures. The machine learning system includes (i) a first subnetwork to generate feature data based on the sensor data, (ii) a second subnetwork to extract a selected patch of the feature data, and (iii) a third subnetwork to generate gesture data based on a classification of the corresponding feature data of the selected patch. The first cross entropy loss is based on a first performance of the second subnetwork in relation to the training dataset. The second cross entropy loss is based on a second performance of third subnetwork in relation to the training dataset.