IMU Activity Recognition with OoD Filtering for Sensor Shift
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Conventional deep learning models for human activity recognition struggle to effectively discriminate between in-distribution and out-of-distribution samples, particularly in open-world environments with varying input conditions, leading to system failures due to distribution shifts and sensor degradations.
Innovation Solution
Implementing an OoD detector using a modified ODIN method and quadruplet loss technique to filter out-of-distribution samples, followed by embedding models to project input features into lower-dimensional vectors for classification, utilizing a classifier unit to handle sensor artifacts and environment uncertainties.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If conventional deep learning models are used for human activity recognition, then the system can process motion data, but it fails to discriminate between in-distribution and out-of-distribution samples leading to system failures
Solution Approach 1:
The patent segments the classification task into two distinct stages: first, an OoD detector identifies and filters out-of-distribution samples; second, a classifier processes only in-distribution samples. This segmentation resolves the contradiction by allowing the system to maintain high reliability for known activities while explicitly handling diverse input conditions through the filtering mechanism.
Solution Approach 2:
The OoD detector acts as an intermediary component between the input motion data and the classifier. It mediates by evaluating whether inputs belong to the training distribution before allowing them to reach the classifier, thus protecting the system from OoD samples that would otherwise cause failures while maintaining normal classification functionality.
2Adaptability or versatility
If the system accepts all input samples for classification, then it can handle diverse environments, but it suffers from distribution shifts and sensor degradations causing system failures
Solution Approach 1:
The system performs preliminary action by evaluating inputs for distribution membership before classification. The OoD detector assesses whether each input sample belongs to the training distribution in advance, filtering out problematic samples (those affected by distribution shifts or sensor degradations) before they can cause system failures, while still allowing diverse valid inputs to proceed.
3Productivity
If the model is trained on limited training data, then it can be deployed quickly, but it cannot reliably handle unknown classes and sensor artifacts in open-world environments
Solution Approach 1:
The patent extracts the OoD detection capability as a separate, independent component that operates alongside the classifier. This extracted OoD detector can be trained on the same limited training data to learn the boundaries of the training distribution, then independently filter out unknown classes and sensor artifacts, allowing the system to deploy quickly while maintaining reliability in open-world environments.
Data Source
AI summary
Implementations disclosed describe methods, devices, and systems to perform out-of-distribution and recognition of activities using an inertial measurement unit (IMU) sensor. A method may include receiving motion data by a device from a motion sensor. The method further includes generating image data comprising one or more images based on the motion data. The method further includes determining that a first portion of the image data corresponds to activities outside a classification distribution. The method further includes filtering the image data by removing the first portion from the image and generating filtered image data. The method further includes determining an activity classification, within the classification distribution, based on the filtered image data. The method further includes modifying an operating parameter of the device based on the activity classification.