System and method for user recognition using motion sensor data
Inventors
Ionescu, Radu Tudor • Ungureanu, Adrian Ionut • Dumitran, Ionut
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-11733780-B2
Publication Date
2023-08-22
Expiration Date
Abstract
Technologies are presented herein in support of system and methods for user recognition using motion sensor data. Embodiments of the present invention concern a system and method for capturing motion sensor data using motion sensors of a mobile device and characterizing the motion sensor data into features for user recognition. The motion sensor data of a user is collected by the motion sensors of a mobile device in the form of a motion signal. One or more sets of features are extracted from the motion signal and a subset of discriminative features are then selected. The subset of features is analyzed, and a classification score is generated to classify the user as a genuine user or an imposter user.
Core Innovation
The invention provides a method for user recognition using motion sensor data. A mobile device having at least one motion sensor collects a motion signal of a user, and the motion signal is analyzed with a plurality of feature extraction algorithms. The analysis extracts sets of features in which each individual set includes discriminative and non-discriminative features extracted by a respective feature extraction algorithm, and the extracted sets are concatenated to form a combined set of extracted features.
The invention performs discriminative feature selection by applying a feature selection algorithm that comprises a principal component analysis algorithm. The principal component analysis algorithm ranks the extracted features based on the level of variability of the feature between users, and selects features with the highest levels of variability to form the subset of discriminative features. This subset of discriminative features is used for user verification as a genuine user or an imposter user.
For classification, the invention uses a classification algorithm comprising a stacked generalization technique that generates a classification score. The stacked generalization technique utilizes more than one classifier selected from Naïve Bayes, Support Vector Machine (SVM), Multi-layer Perception, Random Forest, and Kernel Ridge Regression (KRR). In the system implementation, the stacked generalization technique is organized in two layers, where a first layer provides multiple classifications and a second layer classifies using an output given by the first layer.
Claims Coverage
The patent includes three independent claims, centered on multi-algorithm motion-feature extraction with discriminative and non-discriminative features, PCA-based ranking of features by between-user variability, and stacked generalization classification producing a classification score for genuine versus imposter users.
Multi-algorithm motion feature extraction and concatenation
Extracting sets of features with a plurality of feature extraction algorithms, where each individual set includes discriminative and non-discriminative features, and concatenating the sets of extracted features to form a combined set of extracted features.
PCA-based selection of discriminative features by between-user variability
Using a feature selection algorithm comprising a principal component analysis algorithm to rank extracted features based on the level of variability between users and selecting features with the highest levels of variability to form a subset of discriminative features.
Stacked generalization classification for genuine versus imposter users
Using a classification algorithm comprising a stacked generalization technique with more than one classifier selected from Naïve Bayes, SVM, Multi-layer Perception, Random Forest, and KRR, and generating a classification score to classify the user as genuine or imposter.
HOG with two gradient orientations as part of multi-algorithm extraction
Including a Histogram of Oriented Gradients (HOG) technique that employs two gradient orientations as one of the feature extraction algorithms.
Two-layer stacked generalization in a system for analyzing a motion signal
Organizing the stacked generalization technique in two layers, where the first layer provides multiple classifications and the second layer classifies using an output given by the first layer.
Across the independent claims, the invention combines multi-algorithm extraction of motion-signal feature sets, PCA-based selection by between-user variability, and stacked generalization classification with named classifiers to produce a genuine versus imposter classification score.
Stated Advantages
Documented Applications
No documented applications found
Interested in licensing this patent?