Action Interval Estimation Model With Augmented HSMM Training

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

The high cost of generating supervised data for training models to estimate time segments of actions is a significant challenge.

Innovation Solution

An action segment estimation model is built using a hidden semi-Markov model, where observation probabilities are learned through unsupervised learning, and transition probabilities are learned using supervised learning with augmented supervised data generated by oversampling in time and feature space, incorporating teacher information.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Measurement precision

If supervised data is generated manually for training action segment estimation models, then model training accuracy is improved, but data generation cost and time consumption increase significantly

Engineering Contradiction:
Improvemodel training accuracyVSAvoiddata generation time
Core Design Contradiction:
Measurement precisionVSLoss of time

Solution Approach 1:

The system performs preliminary unsupervised learning to extract movement patterns and templates before the actual supervised training. By pre-processing the data to identify and categorize movement types, the system reduces the manual annotation burden while maintaining training accuracy. The movement templates generated in advance serve as reusable components for subsequent supervised learning.

Inventive Principle:
Principle #10Preliminary action

Solution Approach 2:

The system creates synthetic training data by copying and transforming existing movement templates. Through data augmentation techniques such as temporal scaling, spatial transformation, and noise injection, the system generates diverse training samples from limited annotated data, reducing the need for extensive manual data collection while improving model generalization.

Inventive Principle:
Principle #26Copying

2Measurement precision

If more supervised data is collected for training, then model estimation accuracy is improved, but data collection cost and complexity increase

Engineering Contradiction:
Improveaction segment estimation accuracyVSAvoiddata collection complexity
Core Design Contradiction:
Measurement precisionVSDevice complexity

Solution Approach 1:

The system enables self-service learning by automatically generating movement templates and annotated data from raw video sequences without requiring manual intervention. The unsupervised movement extraction algorithm autonomously identifies action boundaries and generates training labels, allowing the system to improve its performance through self-annotation rather than relying on expensive human annotators.

Inventive Principle:
Principle #25Self-service

Solution Approach 2:

The system changes the parameters of training data by applying various transformations to movement templates, including temporal scaling, speed variation, and spatial transformations. This approach creates diverse training samples from a single annotated example, effectively increasing the training dataset size without additional data collection efforts.

Inventive Principle:
Principle #35Parameter changes

3Manufacturing precision

If manual annotation of action segments is performed, then training data quality is improved, but processing time and labor cost increase

Engineering Contradiction:
Improvetraining data qualityVSAvoiddata processing efficiency
Core Design Contradiction:
Manufacturing precisionVSProductivity

Solution Approach 1:

The system segments the data processing task into two distinct phases: unsupervised movement template extraction and supervised model training. By separating the template extraction phase (which can be automated) from the model training phase (which benefits from quality annotated data), the system achieves both high automation and high data quality. The segmentation allows different processing strategies to be applied to different parts of the pipeline.

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The system introduces movement templates as an intermediary representation between raw video data and the final action segment estimation model. These templates serve as a bridge that captures essential movement patterns without requiring full manual annotation of action segments. The intermediary templates can be generated automatically while still providing sufficient structure for accurate supervised learning.

Inventive Principle:
Principle #24Intermediary (Mediator)

Data Source

PatentEP4287078B1Device for building model for estimating action interval, method for building model for estimating action interval, and program for building model for estimating action interval
Publication Date: 2026.03.11 FUJITSU LTD
  • EP4287078B1 patent drawingFigure 1
  • EP4287078B1 patent drawingFigure 2
  • EP4287078B1 patent drawingFigure 3

AI summary

In a hidden semi-Markov model, observation probabilities for each type of movement of plural first hidden Markov models are learned using unsupervised learning. The hidden semi-Markov model includes plural second hidden Markov models each containing plural of the first hidden Markov models using types of movement of a person as states and with the plural second hidden Markov models each using actions determined by combining plural of the movements as states. The learnt observation probabilities are fixed, input first supervised data is augmented so as to give second supervised data, and transition probabilities of the movements of the first hidden Markov models are learned by supervised learning in which the second supervised data is employed. The learnt observation probabilities and the learnt transition probabilities are used to build the hidden semi-Markov model that is a model for estimating segments of the actions. Augmentation is performed on the first supervised data by adding teacher information of the first supervised data to each item of data generated by at least one out of oversampling in the time direction or oversampling in feature space.