Sensor Data Labeling via Behavior Estimation Models
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing methods for monitoring objects using machine learning-based behavior estimation models often result in low-quality training data due to incorrect labeling or the inability to label sensor data, leading to discarded or mislabeled data.
Innovation Solution
A method and system that acquire labeling target sensor data and determine its labeling information by estimating the object's behavior from a combination of the sensor data and reference data of a different type, thereby minimizing data discard and mislabeling.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If sensor data is labeled based on video data, then labeling can be performed, but the object may be obscured in video data making accurate labeling difficult
Solution Approach 1:
The patent introduces behavior estimation data as an intermediary to bridge the gap between sensor data and labeling. When video data shows obscured objects, the system uses behavior estimation models that process sensor data (acceleration, temperature, humidity) to infer object behavior, which then serves as a mediator for accurate labeling without requiring direct visual observation of the obscured object.
Solution Approach 2:
The patent replaces direct visual inspection (mechanical observation system) with a data-driven behavior estimation system. Instead of relying on visual detection of object status in video data, the system substitutes this with automated behavior estimation models that process multi-type sensor data to determine object behavior, thereby overcoming the limitation of visual obstruction.
2Reliability
If human labelers manually label sensor data, then labeling can be performed, but low judgment proficiency leads to incorrect labeling
Solution Approach 1:
The patent implements self-service labeling where the system automatically labels sensor data using behavior estimation models without requiring human intervention. The behavior estimation model processes sensor data and reference data to automatically determine object behavior and generate labels, eliminating the dependency on human labeler proficiency and ensuring consistent, high-quality labeling.
Solution Approach 2:
The patent incorporates feedback mechanisms where the behavior estimation model continuously refines its labeling based on the relationship between sensor data and reference data. The system uses feedback from the behavior estimation process to improve labeling accuracy, allowing the system to learn from patterns in the data and progressively improve its labeling quality.
3Quantity of substance
If sensor data is collected for training, then training data can be obtained, but data may be discarded or mislabeled reducing training quality
Solution Approach 1:
The patent applies preliminary action by performing behavior estimation and labeling determination before the data is finalised for training. The system preliminarily processes sensor data through behavior estimation models to determine object behavior and generate labels in advance, allowing for quality control and filtering before the data enters the training pipeline, thereby preventing discarded or mislabeled data from reducing training quality.
Data Source
AI summary
According to one aspect of the present invention, provided is a method for supporting labeling of sensor data, comprising the steps of: acquiring sensor data to be labeled, measured by means of a sensor for a subject; and referring to the subject's behavior estimated from the sensor data to be labeled and/or first corresponding reference data, which corresponds to the sensor data to be labeled and belongs to a type differing from that of the sensor data to be labeled, thereby determining information related to labeling of the sensor data to be labeled.