Activity Recognition Model Balancing Versatility and Individuation
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing activity recognition systems face challenges in balancing versatility and individuation, leading to high training costs and compromised accuracy for individual users. Current methods either discard personal features for generalization or require manual intervention and significant data collection for migration.
Innovation Solution
The proposed activity recognition system employs a 'cloud edge' hardware architecture, utilizing edge computing for local data processing and cloud computing for centralized model updates. This system incorporates a model training module that fuses versatile and individualized network structures, enabling balanced recognition through federated learning and minimizing manual intervention.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Adaptability or versatility
If a general activity recognition model is trained using a large amount of user data to improve generalization ability, then the model can be applied to new users without requiring them to provide huge training data, but user-specific features are discarded leading to reduced accuracy for individual activity recognition
Solution Approach 1:
The activity recognition model is segmented into two distinct components: a general model trained on large-scale user data to capture universal activity patterns, and a personalized model trained on individual user data to capture user-specific features. This segmentation allows the system to simultaneously achieve generalization across users and precision for individual recognition, resolving the contradiction between adaptability and measurement precision.
2Measurement precision
If traditional recognition methods are used to maintain accuracy for individual users, then user-specific features are preserved, but new users need to provide a large amount of data to adjust recognition model parameters or re-train a suitable activity recognition model
Solution Approach 1:
The system performs preliminary action by pre-training a general activity recognition model on large-scale user data before deployment. This pre-trained general model serves as a foundation that can be quickly adapted to individual users with minimal training data, eliminating the need for new users to provide large amounts of data or wait for extensive re-training, thus reducing training time while maintaining accuracy.
3Adaptability or versatility
If data-based migration methods are used to reduce training costs for new users, then generalization ability is improved, but considerable network resources are consumed and privacy disclosure risks increase
Solution Approach 1:
The system extracts and transfers only the essential model parameters and feature representations from the general model to the personalized model, rather than migrating entire datasets or complex model structures. This extraction approach reduces network resource consumption during model migration and minimizes privacy risks by avoiding the transfer of sensitive user data, while still achieving effective generalization.
4Measurement precision
If manual intervention is used to determine which parts to migrate in model migration, then migration accuracy can be improved, but migration costs and complexity increase
Solution Approach 1:
The system implements self-service by enabling the model migration process to automatically identify and transfer relevant parameters and features without requiring manual intervention. The framework includes automated mechanisms for determining which model components should be migrated, calculating transfer weights, and integrating source and target models, thereby reducing migration complexity while maintaining high migration accuracy.
Data Source
AI summary
The present invention relates to an activity recognition system balanced between versatility and individuation, comprising a communication framework jointly formed by a data collecting terminal, a computing device, and a cloud computing platform, the activity recognition system uses the communication framework to conduct personnel activity recognition and model updating, and the edge computing device further comprises a model training module and an activity recognition module, and the model training module retrieves a local activity recognition model by continuously verifying user IDs, and uses the first data to train a versatile network structure and an individualized network structure of the local activity recognition model in a way that individuation features of the user and versatility features of the model are fused with each other, so that the personnel activity recognition process conducted by the activity recognition module.