Deep Learning Hand Gesture Recognition in AR

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Conventional wearable Augmented Reality (AR) devices struggle with hand gesture recognition in varying light conditions and reflective surfaces, requiring complex and expensive hardware, and current methods are limited by skin color dependency and lack robustness in realistic conditions.

Innovation Solution

A deep learning-based system for detecting and classifying dynamic hand gestures using a single RGB camera without depth sensors, employing a multi-layer LSTM network to estimate temporal information and classify gestures, overcoming limitations of existing techniques by utilizing RGB image data for robust gesture recognition.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If conventional wearable AR devices use complex on-board sensors and customized processing chips for hand gesture recognition, then gesture recognition accuracy is improved, but device cost and complexity increase

Engineering Contradiction:
Improvegesture recognition accuracyVSAvoidhardware complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent extracts the depth sensing functionality from the AR device hardware and performs gesture recognition using only the standard RGB camera already present in frugal devices. By taking out the requirement for specialized depth sensors and processing chips, the system achieves gesture recognition without increasing device complexity or cost.

Inventive Principle:
Principle #2Taking out (Extraction)

Solution Approach 2:

The system creates a virtual depth map by processing RGB image data through deep learning algorithms. This copied representation of depth information allows the system to achieve accurate gesture recognition without physical depth sensors, resolving the contradiction between accuracy and hardware complexity.

Inventive Principle:
Principle #26Copying

2Reliability

If conventional AR devices use specialized depth sensors for gesture detection, then gesture recognition reliability is improved, but device cost increases

Engineering Contradiction:
Improvegesture recognition reliabilityVSAvoidhardware cost
Core Design Contradiction:
ReliabilityVSDevice complexity

Solution Approach 1:

The patent replaces expensive, specialized depth sensors with a standard RGB camera that is already present in frugal AR devices. This substitution uses cheaper, widely available hardware to achieve reliable gesture recognition, eliminating the need for costly specialized components.

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

Solution Approach 2:

The system changes the input parameter from depth information (requiring specialized sensors) to RGB color information (available from standard cameras). By processing RGB data through multi-layer LSTM networks, the system achieves reliable gesture recognition using parameters already captured by inexpensive hardware.

Inventive Principle:
Principle #35Parameter changes

3Ease of operation

If conventional gesture recognition methods use skin color detection, then detection simplicity is improved, but robustness in varying light conditions deteriorates

Engineering Contradiction:
Improvedetection simplicityVSAvoidrobustness in varying light conditions
Core Design Contradiction:
Ease of operationVSReliability

Solution Approach 1:

The patent introduces a deep learning-based hand pose estimation system as an intermediary between the RGB camera and gesture recognition. This intermediary extracts hand keypoint coordinates through multi-layer LSTM networks, providing a robust representation that is independent of skin color and lighting conditions, thereby resolving the contradiction between simplicity and robustness.

Inventive Principle:
Principle #24Intermediary (Mediator)

4Measurement precision

If conventional systems use multiple cameras for 3D hand pose detection, then measurement precision is improved, but device complexity increases

Engineering Contradiction:
Improve3D hand pose detection accuracyVSAvoidcamera system complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent makes the single RGB camera perform multiple functions: capturing images for display, extracting hand keypoints for gesture recognition, and inferring 3D hand pose information. This multi-functionality allows the system to achieve 3D pose detection accuracy without adding multiple cameras, resolving the contradiction between precision and complexity.

Inventive Principle:
Principle #6Universality (Multi-functionality)

Solution Approach 2:

The system creates virtual 3D hand pose representations by processing 2D RGB image data through deep learning algorithms. This copied 3D information from 2D images enables accurate hand pose detection without physical multi-camera systems, eliminating the need for complex hardware while maintaining measurement precision.

Inventive Principle:
Principle #26Copying

Data Source

PatentEP3467707B1System and method for deep learning based hand gesture recognition in first person view
Publication Date: 2024.03.13 TATA CONSULTANCY SERVICES LTD
  • EP3467707B1 patent drawingFigure 1A~1D
  • EP3467707B1 patent drawingFigure 2~3
  • EP3467707B1 patent drawingFigure 4

AI summary

A system and method for hand-gesture recognition are provided. The method includes receiving frames of a media stream of a scene captured from a first person view (FPV) of a user using at least one RGB sensor communicably coupled to a wearable AR device. The media stream includes RGB image data associated with the frames of the scene. The scene comprises a dynamic hand gesture performed by the user. A temporal information associated with the dynamic hand gesture is estimated from the RGB image data by using a deep learning model. The estimated temporal information is associated with hand poses of the user and comprising a plurality of key-points identified on user's hand in the plurality of frames. Based on the temporal information of the key points, the dynamic hand gesture is classified into at least one predefined gesture class by using a multi-layered LSTM classification network.