TL;DR
This paper introduces a novel zero-velocity update (ZUPT) algorithm for foot-mounted pedestrian navigation systems (PNS) that significantly improves accuracy. By analyzing speed variations over a gait cycle using a finite state machine, their ZUPT method achieves a mean relative accuracy of 0.89% across various motion modes, outperforming existing methods.
What Problem Does This Paper Address?
Accurately tracking pedestrian movement, especially indoors or in GPS-denied environments, is a persistent challenge. Foot-mounted inertial sensor-based systems (PNS) offer a promising solution due to their low cost and small size. However, these systems suffer from rapid error accumulation, making position errors grow cubically with time. To combat this, Zero-Velocity Updates (ZUPT) are crucial. ZUPT leverages the fact that a pedestrian's foot is momentarily stationary during each gait cycle to reset velocity errors and correct sensor biases. The core problem this paper addresses is the reliable and robust detection of these zero-velocity intervals across diverse human gaits (walking, running, stair climbing, etc.), as existing methods often struggle with variations in motion modes or are prone to false positives.
Methods
The paper proposes a novel ZUPT algorithm built upon a Kalman filter framework. Key methodological contributions include:
- Error Model Development: A detailed error model for the foot-mounted PNS is established, forming the basis for the Kalman filter.
- Novel Zero Velocity Detection: Instead of relying on acceleration or angular rate thresholds (which vary significantly with motion), this method focuses on the variations in speed over a gait cycle.
- Finite State Machine (FSM): A three-state FSM is introduced to model a gait cycle:
- State 1: Zero velocity interval.
- State 2: Acceleration interval.
- State 3: Deceleration interval.
- Speed-Based State Transitions: State transitions are determined using a sliding window approach that analyzes the speed profile. Specifically, the slope of the speed (representing acceleration) and the maximum/minimum speeds within the window trigger transitions between states. This approach is designed to be more robust to different motion types.
- Kalman Filter Integration: The detected zero-velocity moments are used as measurements in a Kalman filter to estimate and correct sensor biases (gyroscope and accelerometer), attitude errors, and velocity errors.
Datasets
The paper describes experiments conducted using a pedestrian and an Inertial Measurement Unit (IMU). The IMU used was a model MTI-G from Xsens Technologies, providing data at 100 Hz. The experiments involved a 25-year-old male pedestrian performing six types of motion: walking, running, stair climbing, stair descending, uphill, and downhill. Additional tests were conducted with different IMU mounting positions (front vs. rear of the shoe) and with different pedestrians and shoe types to assess robustness. No specific public datasets are mentioned; the evaluation is based on custom experimental data.
Key Insights and Ideas
- Speed Variation is a More Robust Indicator: The paper's central insight is that analyzing the speed profile of a gait cycle is more effective for detecting zero-velocity intervals than relying on acceleration or angular rate. This is because speed variations tend to be more consistent across different motion modes compared to the highly variable acceleration and angular rate signals.
- Finite State Machine for Gait Cycle Modeling: Using a finite state machine to explicitly model the phases of a gait cycle (acceleration, deceleration, zero velocity) provides a structured and interpretable way to process the speed data.
- Insensitivity to Mounting Position: A significant finding is that the proposed algorithm's performance is largely insensitive to where the IMU is mounted on the shoe (front or rear), suggesting flexibility in practical deployment.
- Robustness Across Motion Modes: The algorithm demonstrates strong performance across a wide range of activities, including walking, running, and stair climbing, which is a key limitation of many prior ZUPT methods.
Key Results
- Mean Relative Accuracy: The proposed algorithm achieved a mean relative accuracy of 0.89% across various motion modes.
- Comparison to Prior Work: The paper compares its method to algorithm [11] and shows superior performance. For instance, in running, algorithm [11] missed most zero-velocity intervals, while the proposed method detected them correctly. Similarly, for stair climbing and descending, algorithm [11] performed poorly or incorrectly.
- Trajectory Accuracy: In field tests, the proposed algorithm achieved a mean error of approximately 1.10% of the travelled distance (4.64 m over 422 m) for a mixed walking/running trajectory, significantly better than algorithm [11]'s 3.75% (15.82 m). For a more complex uphill/stair descending trajectory, the mean error was even lower at 0.68%.
- Robustness: The algorithm performed well with different pedestrians and shoe types, indicating good generalization.
Limitations and Future Work
- Elevator Environment: The paper notes that in an elevator environment, the algorithm struggled because the acceleration was below the threshold ($Th_a$), causing the ZUPT state machine to remain in the zero-velocity state continuously. This is identified as a problem to be solved in the future.
- Heading Error: The current ZUPT approach cannot estimate or compensate for heading errors, which are a primary source of error in PNS.
- Future Integration: The authors suggest integrating their ZUPT technique with other methods like Received Signal Strength (RSS) measurements, map matching, building heading information, or computer vision for further accuracy improvements. This integration is planned for future work.
Summary updated