Activity Recognition Model Bias Correction via Class Equalization
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Existing human activity recognition systems tend to be biased towards the most frequent classes due to skewed sample distributions in training and test data sets, leading to poor performance in recognizing less represented activities.
Innovation Solution
The system equalizes the number of samples for each activity class by truncating sequences to a common length, and employs a convolutional neural network (CNN) with a softmax layer and a Finite State Machine (FSM) for accurate classification, using data from accelerometers and gyroscopes to recognize activities without bias.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Device complexity
If simple models are employed to determine activity based on accelerometer data, then device complexity is reduced, but measurement precision and reliability of activity recognition deteriorate
Solution Approach 1:
The patent combines multiple sensor types (accelerometer, gyroscope, magnetometer) and multiple data processing approaches (deep learning neural networks, traditional machine learning classifiers) to create a hybrid system that achieves high recognition accuracy while managing complexity through integrated processing
2Ease of manufacture
If training data contains skewed class distribution, then data collection is simpler, but recognition reliability for less frequent activities deteriorates
Solution Approach 1:
The patent applies data augmentation techniques during the training phase to artificially balance the class distribution before model training. This preliminary action ensures that minority classes are adequately represented, improving recognition reliability without requiring complex data collection procedures
Solution Approach 2:
The patent employs class weighting parameters in the loss function to adjust the importance of different classes during training. By modifying this parameter, the system gives higher weight to minority classes, ensuring they are properly learned despite their lower frequency in the original dataset
3Manufacturing precision
If sequence truncation is applied to equalize sample numbers, then class distribution becomes balanced, but information loss occurs
Solution Approach 1:
Instead of truncating sequences in the time dimension, the patent uses data augmentation to add synthetic samples in the feature dimension. This approach balances class distribution without reducing the temporal information content of original sequences
4Measurement precision
If deep learning models are used for activity recognition, then measurement precision improves, but device complexity and computational requirements increase
Solution Approach 1:
The patent segments the processing architecture into distinct components: a deep learning-based activity recognition module and a separate Finite State Machine module for temporal consistency. This segmentation allows each component to specialize, maintaining high accuracy while managing overall system complexity
5Reliability
If Finite State Machine is added for prediction stability, then reliability improves, but device complexity increases
Solution Approach 1:
The Finite State Machine acts as an intermediary layer between the deep learning classifier and the final output. It receives probabilistic predictions from the neural network and applies temporal logic rules to produce stable, consistent activity labels, mediating between raw model output and reliable recognition results
Data Source
Figure 1
Figure 2
Figure 3
AI summary
Human activities are classified (152, 158) based on activity-related data and an activity-classification model trained using a classification-equalized training data set. A classification signal is generated (158) based on the classifications. The classification-equalized training data set, may, for example, includes a first class having a first sequence length and a number of samples N, and one or more additional classes each having a respective sequence length tj and a respective number of samples Nj determined based on the number of samples N of the first class. For example, a respective sequence length tj and a respective number of samples Nj which satisfy: (i) Nj > N, for sequence length tj; and (ii) Nj < N, for tj - 1. The activity-related data may include one or more of acceleration data (122), orientation data (160), position data (126), and physiological data (124).