TL;DR
This paper introduces a novel hybrid clustering model for accurate, real-time recognition of multiple racquet sports movements using a single wearable sensor. The system combines a multi-layer K-means and DBSCAN approach to effectively distinguish similar movements and count repetitions, offering a practical solution for fitness tracking and sports analysis.
What Problem Does This Paper Address?
Physical activity recognition (PAR) is a growing field with significant implications for healthcare, fitness, and sports analysis. While various methods exist for PAR, accurately distinguishing between similar movements, especially in dynamic activities like racquet sports, remains a challenge. Traditional vision-based methods often require specialized equipment and are susceptible to environmental factors. Existing wearable sensor-based approaches can struggle with the subtle differences between strokes within the same sport or across different racquet sports. This paper aims to bridge this gap by developing a system that can reliably identify and count multiple racquet sports movements using a single, unobtrusive wearable sensor, enabling real-time feedback and data sharing.
Methods
The core of this work is a multilayer hybrid clustering model designed for feature extraction and classification. The model operates in four layers:
- Layer 1 & 2 (K-means Clustering): These layers use K-means to segment and cluster extracted features (ASMV, VSMV, DSMV, OSMV) from the sensor data. The first layer focuses on feature maxima from sliding windows, while the second layer clusters these into sub-features, sorting them by frequency to aid in distinguishing commonalities and unique aspects of movements.
- Layer 3 (K-means Clustering): This layer further refines the sub-features by segmenting them with a sliding window and applying K-means to identify sub-feature centers.
- Layer 4 (DBSCAN Clustering): The final layer employs DBSCAN to cluster these sub-feature centers. This helps identify common movement features and outliers, effectively distinguishing between different racquet sports movements even in the presence of noise or similar sub-features. A greedy algorithm is used to optimize the feature sets.
The system also incorporates a wearable wristband equipped with an IMU (MPU6050), an MCU (STM32F103), and Bluetooth (CC2541) for data collection, real-time processing, and communication. Data preprocessing includes median filtering and the use of a complementary filter (HBL) for noise reduction.
Datasets
The paper describes a data collection process involving 5 healthy subjects (3 male, 2 female) with varying levels of experience in badminton and table tennis. Participants performed nine types of movements: four for table tennis (service, stroke, spin, picking up), four for badminton (service, drive, smash, picking up), and walking. For each subject, 20 tests per movement were conducted, resulting in approximately 900 instances per movement set. The data was collected in a real training environment. For model evaluation, three subjects' data (60 instances) were used for training, and the other two subjects' data (40 instances) were used for testing, employing a fivefold cross-validation. No specific public dataset was mentioned as being used.
Key Insights and Ideas
- Hybrid Clustering for Subtle Differences: The combination of K-means for initial feature segmentation and DBSCAN for outlier detection and robust clustering proves effective in differentiating movements with very similar characteristics, which is a common challenge in racquet sports.
- Wearable Sensor Sufficiency: The study demonstrates that a single wearable sensor on the wrist is sufficient for recognizing a range of racquet sports movements, reducing the need for complex multi-sensor setups or external equipment.
- Real-time Processing on MCU: The system's ability to perform complex feature extraction and classification on a low-power MCU in real-time is a significant achievement, enabling immediate feedback to the user.
- Unlabeled Data Handling: The model's capability to identify unlabeled and noisy data without explicit calibration is crucial for practical applications where perfect data labeling is often difficult to obtain.
Key Results
- The proposed multilayer hybrid clustering model achieved an average recognition accuracy of 86.32% with an F1 score of 82.98% across nine movements.
- For single racquet sports, the accuracy improved: table tennis recognition reached 92.51% (F1 score 92.73%), and badminton recognition reached 94.69% (F1 score 94.53%).
- The wristband verification showed an average accuracy of 77% in real-time tests.
- The system's recognition time for different movements averaged around 1 second, with Bluetooth transmission to the mobile phone taking approximately 0.5 seconds.
- Compared to other classifiers like LDF, Random Forests, and SVM, the proposed model achieved the best accuracy (0.86) while having a competitive training time (27.73s) compared to SVM (145.54s).
Limitations and Future Work
- Dataset Specificity: The current dataset is primarily composed of target movements. The authors acknowledge that in broader scenarios (e.g., daily activities), misidentification might occur due to similar sub-features.
- Accuracy Variation: While overall accuracy is good, the model's precision for certain hitting movements (e.g., drive, stroke) is not as high due to their subtle differences.
- Future Work: The authors suggest collecting more types of movement data to improve overall accuracy and to evaluate more detailed movement aspects. They also aim to further refine the model for wider usage scenarios beyond just racquet sports.
Summary updated