Continuous Self-Learning Person Recognition System
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Traditional person recognition systems fail when the sensor cannot obtain information related to the characteristics included in the training data, such as when a person presents a back view, and they discard data if no training data exists for a person, leading to incomplete recognition and loss of valuable information.
Innovation Solution
The system employs continuous self-learning by collecting and saving additional data in the engagement zone, even if it cannot be used for immediate recognition, and maintains this data locally to associate with global training data later, allowing for improved recognition over time and using cues like clothing and surroundings for identification.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If traditional face-detection-based solutions use offline deep learning techniques to create models from significant amounts of data, then the recognition accuracy is improved when training data is available, but the system fails when the sensor cannot obtain information related to characteristics included in the training data (e.g., back view, obscured face)
Solution Approach 1:
The system transitions from static offline deep learning models to dynamic online learning that continuously adapts in real-time. The recognition system dynamically adjusts its parameters and learns new characteristics during operation, enabling it to handle varying conditions such as back views, obscured faces, and changing environments without requiring complete retraining.
Solution Approach 2:
The system performs preliminary data collection and local learning before global model updates. By collecting data locally and performing preliminary analysis, the system prepares training data in advance, which is then used to update the global model. This preliminary action ensures that the system has ready-to-use data for improving recognition accuracy under various conditions.
2Manufacturing precision
If the system discards data when no training data exists for a person, then the system maintains data quality standards, but valuable information is lost that could improve future recognition
Solution Approach 1:
Instead of permanently discarding data that doesn't match existing training data, the system recovers and stores this data locally for future use. The local learning component captures and retains data about persons without established training data, which can then be used to create new training data or improve existing models, thus recovering potentially valuable information.
Solution Approach 2:
The system performs self-service by automatically collecting, storing, and utilizing data from unknown persons without requiring manual intervention. The local learning mechanism autonomously captures data and uses it for future recognition improvements, enabling the system to serve itself by expanding its own training data capabilities.
3Reliability
If the system relies solely on initial training data, then the system operates with a clear recognition baseline, but the system cannot adapt to new conditions or improve recognition accuracy over time
Solution Approach 1:
The system implements feedback loops where recognition results and collected data continuously inform model updates. The online learning component uses feedback from real-time data collection and recognition attempts to continuously improve the global model, enabling the system to adapt to new conditions while maintaining recognition baseline through structured model updates.
Solution Approach 2:
The system maintains continuous learning and adaptation through online data collection and model updates. Rather than stopping at initial training, the system continuously performs useful actions by collecting data, updating models, and improving recognition capabilities over time, ensuring both baseline reliability and ongoing adaptability.
Data Source
AI summary
A system and related method are provided for sensor-based person recognition. The system has a data store including global data of a user, the global data including a globally unique identifier and sensed training data. The system further has a classifier to identify the user at a first time using the global data, augment the global data with first context information of the user captured at the first time, and identify the user at a second time by finding correlations between second context information and the first context information when using the global data fails, the second context information of the user captured at the second time.