TL;DR
This paper introduces a novel gait recognition system using optical motion capture (OMCS) that achieves high accuracy by combining frame-level classification with a new "Reliability-Weighted Sum" (RWS) decision fusion method. This approach allows for accurate identification even with a limited number of motion frames and trackers, potentially enabling more efficient and privacy-preserving surveillance and authentication.
What Problem Does This Paper Address?
Traditional human identification methods like facial recognition or fingerprinting raise privacy concerns. Gait recognition offers a privacy-friendly alternative, identifying individuals by their unique walking patterns. While optical motion capture systems (OMCS) can capture highly precise 3D human motion data, their application in gait recognition, especially in scenarios requiring simultaneous tracking of multiple individuals, has been underexplored. Existing gait recognition methods often require extensive training data (multiple gait cycles per person) and may struggle with limited data or noisy sensor readings. This paper aims to bridge this gap by developing an OMCS-based gait recognition system that is efficient, accurate even with limited data, and capable of handling multiple subjects.
Methods
The proposed method follows a four-step pipeline:
- Feature Extraction: Instead of using raw motion data, the system extracts features from each individual motion frame. This involves calculating the relative Euclidean distances between 10 optical trackers placed on the lower body (thigh, lower leg, foot) and also incorporates the x-axis velocity of each tracker. This approach aims to capture both body shape and motion characteristics.
- Unreliable Feature Calibration: Recognizing that sensor noise or failures can lead to biased motion data, the paper proposes a method to detect and calibrate unreliable features. This is done by estimating the probability density function (PDF) of each feature using kernel density estimation and identifying outliers based on probability thresholds. These outliers are then adjusted to mitigate their impact on classification.
- Single Motion Frame Classification: For each calibrated motion frame, a Kernel Extreme Learning Machine (KELM) classifier is used. KELM is chosen for its superior classification accuracy and efficiency compared to traditional methods like Support Vector Machines (SVM) and Random Forests, especially when dealing with limited training data.
- Decision Fusion of Multiple Motion Frames: To improve overall accuracy and robustness, the decisions from individual frames are combined. The paper introduces a novel Reliability-Weighted Sum (RWS) rule. This method first transforms the KELM classifier's continuous outputs into fuzzy decisions. Then, it calculates a consistency matrix between these fuzzy decisions. Using eigenvalue decomposition, it derives reliability weights for each fuzzy decision, giving higher weight to decisions that are more consistent with others. Finally, these weighted fuzzy decisions are combined to produce a global decision.
Datasets
The study utilizes a dataset collected from 76 participants (46 female, 30 male) with varying ages and heights. The participants wore 10 optical trackers on their lower body. Walking data was collected at a sampling frequency of 5 Hz. The dataset was split into training and testing sets, with the training set containing approximately 9595 motion frames and the testing set containing approximately 4787 motion frames. The paper does not mention the introduction of a new benchmark dataset.
Key Insights and Ideas
- OMCS for Gait Recognition is Underexplored: The paper highlights that while OMCS is widely used in other fields, its potential for gait recognition, especially for multi-person scenarios, has not been thoroughly investigated.
- KELM's Superiority for Limited Data: The research demonstrates that KELM significantly outperforms SVM and Random Forest for single-frame gait classification, particularly when dealing with limited training data, which is common in OMCS-based gait analysis.
- The Power of Decision Fusion: The core innovation lies in the RWS decision fusion method. The insight is that not all frame-level decisions are equally reliable. By quantifying and leveraging the consistency of fuzzy decisions, the RWS rule can effectively filter out unreliable information and improve overall classification accuracy.
- Lower Body Trackers are Sufficient: The study shows that using only 10 trackers on the lower body is sufficient to achieve very high classification accuracy, reducing the intrusiveness and cost associated with full-body tracking.
- Calibration Matters: Calibrating unreliable features, even if they are not deleted, leads to improved classification performance, suggesting that robust handling of noisy sensor data is crucial.
Key Results
- Single Frame Classification: KELM achieved significantly higher accuracy than SVM and Random Forest. With 10 motion trackers, KELM achieved 83.55% accuracy, compared to 56.17% for SVM and 50.44% for Random Forest. Even with 6 trackers, KELM reached 68.17% accuracy.
- Decision Fusion Accuracy: The proposed RWS decision fusion rule consistently outperformed other conventional fusion rules (majority voting, sum rule, product rule, belief rule, reliability-weighted belief rule).
- 100% Accuracy Achieved: With 10 motion trackers and a sufficient number of gait motion frames (less than 50 frames in some tests), the proposed method achieved 100% validation accuracy.
- Efficiency: KELM demonstrated significantly faster training times compared to SVM and Random Forest.
Limitations and Future Work
- Implementation Cost: The paper acknowledges that the high implementation cost of optical trackers and capture cameras limits the widespread application of this specific OMCS-based system.
- Multi-Player Tracking Complexity: While the paper aims to enable multi-player tracking, the current focus is on the recognition method itself. Future work will involve applying the system to more complex multi-player tracking scenarios.
- Complex Trajectories: The current study primarily uses walking gait. Future work will explore the system's performance with more complex trajectories.
- Integration with Other Sensors: The authors suggest that integrating range and depth sensors with video surveillance could improve existing systems, hinting at potential future research directions beyond OMCS.
Summary updated