Virtual Optical Flow Feature Generation for Still Image Action Recognition

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Action recognition in still images is challenging due to limited information, leading to lower accuracy compared to video analysis, as existing methods primarily rely on spatial features without effectively incorporating time attributes.

Innovation Solution

The method generates a virtual optical flow feature for a still image by combining spatial features with optical flow features from a preset library, using weighted summation based on similarities, to simulate time sequence information and improve recognition accuracy without requiring a training model.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If only spatial features are used for action recognition in still images, then the method is simple, but the recognition accuracy is low due to lack of time attribute information

Engineering Contradiction:
Improveaction recognition accuracyVSAvoidfeature processing complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The patent pre-trains a convolutional neural network model on video data to extract spatial and optical flow features, storing them in a feature library before actual action recognition. This preliminary preparation enables the system to efficiently generate virtual optical flow features for still images without requiring complex real-time processing, thereby improving recognition accuracy while controlling computational complexity

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The patent introduces virtual optical flow features as an intermediary representation that bridges spatial features and temporal information. By generating virtual optical flow features through weighted combination of real optical flow features from the feature library, the system effectively incorporates time attributes into still image analysis without directly processing complex temporal data, thus improving accuracy while managing complexity

Inventive Principle:
Principle #24Intermediary (Mediator)

2Measurement precision

If a large quantity of optical flow features are used, then the recognition accuracy improves, but the computational complexity and processing time increase

Engineering Contradiction:
Improveaction recognition accuracyVSAvoidprocessing speed
Core Design Contradiction:
Measurement precisionVSProductivity

Solution Approach 1:

The patent selects a subset of K optical flow features from the feature library that are most relevant to the input spatial features, rather than using all available features. This partial action approach maintains recognition accuracy by focusing on the most informative features while significantly reducing computational complexity and processing time compared to using the complete feature set

Inventive Principle:
Principle #16Partial or excessive action

Solution Approach 2:

The patent dynamically adjusts the quantity and selection of optical flow features based on the specific input image and task requirements. By changing the parameter of feature quantity from a fixed large number to a variable subset size K, the system optimizes the balance between recognition accuracy and processing speed for different scenarios

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP3757874B1Action recognition method and apparatus
Publication Date: 2023.10.25 HUAWEI TECH CO LTD
  • EP3757874B1 patent drawingFigure 1
  • EP3757874B1 patent drawingFigure 2
  • EP3757874B1 patent drawingFigure 3

AI summary

This application relates to artificial intelligence, and provides an action recognition method and apparatus. The method includes: obtaining a to-be-processed picture; extracting a spatial feature of the to-be-processed picture; determining a virtual optical flow feature of the to-be-processed picture based on the spatial feature of the to-be-processed picture and X spatial features and X optical flow features in a preset feature library, where there is a one-to-one correspondence between the X spatial features and the X optical flow features in the preset feature library, and X is an integer greater than 1; determining a first type of confidence of the to-be-processed picture in different action categories based on similarities between the virtual optical flow feature of the to-be-processed picture and Y optical flow features in the feature library, where each of the Y optical flow features in the preset feature library corresponds to one action category, and Y is an integer greater than 1; and determining an action category of the to-be-processed picture based on the first type of confidence. This application can improve action recognition accuracy.