TL;DR
This paper introduces a novel target tracking strategy for wireless sensor networks (WSNs) that combines Support Vector Machines (SVM) for initial position estimation with an improved Kalman Filter (KF) for refined tracking. This approach aims to overcome the limitations of traditional RSSI-based tracking, offering enhanced accuracy and stability, particularly in challenging non-Gaussian noise environments.
What Problem Does This Paper Address?
Accurate target tracking (TT) is a crucial application for wireless sensor networks (WSNs). While Received Signal Strength Indication (RSSI) is an attractive metric for TT due to its ubiquity in commercial devices and minimal hardware requirements, RSSI-based methods often suffer from low stability and precision. This is due to factors like multipath propagation, occlusions, and sensor calibration issues, which lead to noisy and nonlinear measurements. Existing methods, including traditional Kalman Filters, struggle with these nonlinearities and noise, leading to degraded tracking performance. This paper addresses the need for a more robust and accurate RSSI-based TT approach that can handle these real-world challenges.
Methods
The proposed method, SVM+KF, consists of two main phases:
- Offline SVM Model Training:
- Data Preprocessing: The collected RSSI data and corresponding object positions are preprocessed using DBSCAN (Density-Based Spatial Clustering of Applications with Noise) to remove outliers and improve data quality.
- SVM Model Training: A Support Vector Machine (SVM) is trained to learn the nonlinear relationship between RSSI values and target positions. The Radial Basis Function (RBF) kernel is chosen for its suitability with nonlinear and non-Gaussian RSSI data. This phase establishes an initial "prelocalization" model.
- Online Target Tracking:
- Initial Estimation: The trained SVM model is used with incoming RSSI measurements to obtain an initial estimate of the target's position.
- Improved Kalman Filter: An enhanced Kalman Filter (KF) is then employed to refine this initial estimate. The improvement lies in modifying the innovation update step. A threshold value, $E_{Threshold}$, is introduced, which dynamically adjusts based on target speed, sampling frequency, and node communication range. This threshold helps to limit the impact of RSSI outliers and improve tracking stability, especially in dynamic scenarios.
Datasets
The paper uses simulated data for training and evaluation. The RSSI values for simulations were generated using the Okumura-Hata model, incorporating various levels of measurement noise and different motion models (Constant Velocity - CV, and Constant Acceleration - CA). The authors also conducted real-world experiments using a WSN deployed at Nanjing Normal University, collecting 800 RSSI signals for offline training and evaluating the algorithm's performance.
Key Insights and Ideas
- SVM for Nonlinearity: SVM proves effective in capturing the complex, nonlinear relationship between RSSI and target position, a significant improvement over linear models.
- Improved KF for Stability: The innovation threshold in the KF is a critical component for enhancing tracking stability. By dynamically adjusting this threshold, the filter can better handle noisy RSSI measurements and prevent erroneous updates.
- Synergy of SVM and KF: The combination of SVM's prelocalization capability and the KF's recursive refinement offers a powerful approach. SVM provides a good initial guess, and the improved KF then smooths and corrects it, leading to superior overall performance.
- Robustness to Noise: The proposed method demonstrates good performance even with non-Gaussian noise, a common issue in WSNs that degrades traditional methods.
Key Results
- Superior Accuracy: Simulations show that the SVM+KF algorithm achieves the highest tracking accuracy, with mean RMSE values significantly lower than traditional KF, improved KF, PF, RR+KF, and GRNN+KF across various motion models and scenarios. For instance, in simulations, SVM+KF was 3.126% to 12.287% lower in RMSE than other algorithms.
- Enhanced Stability: The variance of localization error for SVM+KF was consistently the lowest among the compared algorithms, indicating superior stability, particularly in challenging turning regions.
- Real-World Validation: Experimental results corroborated the simulation findings, showing that SVM+KF achieved the lowest RMSE (0.71 m) and variance (0.57) in the entire trajectory and the turning region, outperforming RR+KF and GRNN+KF.
- Scalability: The RMSE generally decreases as the number of anchor nodes increases, and SVM+KF maintains its superior performance across different node distributions.
Limitations and Future Work
- Motion Model Dependence: While the SVM+KF performs well with standard CV and CA models, tracking in turning regions remains a challenge. The authors suggest that a better motion model could further improve performance.
- Computational Complexity: Although the running time of SVM+KF is comparable to other LBE methods, it is slightly more complex due to the dynamic setting of the innovation threshold.
- Future Work: The authors suggest extending the approach to incorporate other filtering methods like EKF, UKF, and PF, which could potentially lead to further gains in tracking accuracy. They also mention that while this study focused on combining SVM with KF, other filtering methods could be explored.
Summary updated