Abstract
Recently, vision-based hand gesture recognition techniques have been developed for assisting elderly and disabled people to control home appliances. Frequently occurred problems which lower the hand gesture recognition rate are due to the inter-person variation and intra-person variation. The recognition difficulty caused by inter-person variation can be handled by using user dependent model and model selection technique. And the recognition difficulty caused by intra-person variation can be handled by using fuzzy logic. In this paper, we propose multivariate fuzzy decision tree learning and classification method for a hand motion recognition system for multiple users. When a user starts to use the system, the most appropriate recognition model is selected and used for the user.
Introduction
Vision-based gesture recognition is a promising interface for multi-user environments, particularly for assisting those with limited mobility in controlling smart home appliances. However, real-world deployment faces two primary hurdles: inter-person variation, where different users perform the same gesture differently, and intra-person variation, where a single user's movements fluctuate due to fatigue, mood, or posture. Traditional "hard" decision boundaries often fail to capture this inherent human "fuzziness." This paper introduces a personalized system that leverages Soft Computing—specifically Multivariate Fuzzy Decision Trees (MFDT)—to create robust, user-specific recognition models without requiring extensive initial training for every new user.
Proposed Methodology
The system architecture follows a structured pipeline from raw video input to appliance control. The process begins with segmenting face and hand regions from the video stream. Using stereo vision, the system acquires a 3D trajectory of the hand. Velocity information is then utilized to precisely extract the start and end points of a gesture, ensuring that only intentional movements are processed for feature extraction.
Multivariate Fuzzy Decision Trees (MFDT)
At the core of the classification engine is the MFDT. Unlike standard decision trees that split data based on a single attribute (univariate splits), MFDT utilizes a vector of attributes to create multivariate splits. The system employs Linear Discriminant Analysis (LDA) to determine the optimal attribute vector that maximizes information gain at each node.
As shown in the figure above, multivariate splits allow for diagonal decision boundaries that better fit the natural distribution of gesture data compared to the "staircase" boundaries of univariate trees. By integrating Fuzzy Logic, the system establishes "soft" boundaries (membership functions) instead of rigid thresholds. This approach allows the model to handle the natural noise and variation in human movement more gracefully than traditional methods.
Model Selection and Garbage Rejection
To achieve personalization without long calibration phases, the system maintains a "Model Pool" of pre-trained MFDTs. Upon detecting a new user, a Maximum Likelihood Model Comparison is performed on a small sample of gestures to select the best-fitting model from the pool.
Furthermore, to prevent accidental triggers in a home environment, the authors implemented a state automata-based "Garbage Motion Rejection" layer. This layer filters out common daily activities—such as drinking water, reading a book, or eating snacks—that might otherwise be misclassified as gestures.
Experimental Evaluation
The researchers validated the MFDT algorithm using UCI benchmark datasets (Iris and Wine) and a custom hand gesture dataset involving 10 users.
Dataset Details: * Gestures: 10 distinct motions including Up, Down, Left, Right, Forward, Backward, Circles, and Semicircles. * Features: 16 features extracted from 3D trajectories, including path length, min/max coordinates, and timing. * Garbage Data: 5 common daily activities recorded to test rejection capabilities.
Key Results
- Algorithm Robustness: MFDT outperformed standard C4.5 and C5.0 algorithms on UCI benchmarks, achieving 96.5% accuracy on the Iris dataset and 89.9% on the Wine dataset.
- Gesture Accuracy: The personalized system reached an average recognition accuracy of 93.4% across all users.
- False Positive Mitigation: The system successfully rejected 83.9% of "garbage" motions, demonstrating high reliability for real-world home automation.
Conclusion
By embracing "fuzziness" as a feature rather than a limitation, this research demonstrates that personalized gesture recognition can be both accurate and efficient. The combination of LDA-based multivariate splits and a pre-trained model pool provides a viable path for deploying assistive technologies in diverse, multi-user household environments. Future work will focus on expanding the diversity of the model pool and adapting these principles to modern high-dimensional sensing technologies.
Summary updated