- The paper demonstrates that neural networks can abstract complex sensor fusion tasks in pose estimation without heavy biomechanical models.
- The methodology compared data-driven, hybrid, and hybrid classical approaches, showing neural networks effectively handle magnetic disturbances.
- Empirical results indicate that combining classical filters with neural networks achieves competitive accuracy for both low-cost and high-end MoCap systems.
Deep Inertial Pose: A deep learning approach for human pose estimation
The paper "Deep Inertial Pose: A deep learning approach for human pose estimation" presents a comparative analysis of neural network architectures to estimate human pose using both low-cost and high-end Motion Capture (MoCap) systems equipped with Magnetic, Angular Rate, and Gravity (MARG) sensors. The authors discuss the implementation of neural networks to abstract from complex biomechanical models and analytical mathematics traditionally required for pose estimation.
Overview
Human Pose Estimation (HPE) through inertial data presents unique challenges compared to other methods, such as image-based tracking. The integration of data from MARG sensors requires sophisticated sensor fusion techniques to mitigate individual sensor limitations, like gyroscope data drift or magnetic disturbances affecting magnetometer outputs. Traditionally, sensor fusion has relied on model-based approaches like Kalman Filters (KF) or Madgwick's Gradient Descent filter. This paper investigates an alternative: leveraging neural networks to synthesize sensor data without explicitly modeling the complexities of human motion.
Methodology
The authors developed a multi-step pipeline. This included sensor calibration, aligning sensor data with respective body segments, followed by sensor fusion using various architectures. Importantly, the pipeline aimed to assess whether neural networks could improve pose estimation, particularly in scenarios involving magnetic interferences. Both low-cost sensors within an environment full of interference and high-quality sensors in optimal conditions were tested.
Three broad categories of sensor fusion methods were evaluated:
- Model-Free Neural Networks:
- These rely solely on data-driven approaches and represented a baseline for evaluating the novelty of introducing neural networks into inertial pose estimation.
- Hybrid-Complementary Neural Networks:
- These integrate gyroscope data with neural network predictions, aiming to combine explicit filtering with learned compensation for disturbances.
- Hybrid Classical Filter-Neural Networks:
- This approach involves applying classical filters followed by neural network-based corrections to refine orientation estimations further.
Results and Discussion
The comparative analysis showed that combining classical sensor fusion filters with neural network corrections yielded significant promise. On high-end systems, certain hybrid architectures offered errors comparable to the best-performing classical filters without neural components, positioning neural networks as effective complements rather than replacements.
The paper's ablation studies further reinforced several conclusions:
- Use of Magnetometer: Neural networks can handle magnetic disturbances better than classical filters alone, emphasizing the utility of neural networks in environments prone to such interferences.
- Output Representation: Quaternions remain a robust choice for orientation representation, maintaining lower errors in general.
- Windowed Training: Larger temporal windows during training improved model performance in non-complementary neural network approaches, but smaller windows led to accumulated integration errors.
The neural networks demonstrated an ability to abstract complex underlying models, suggesting that they may offer scalable solutions for real-time HPE across diverse applications—from rehabilitation to sports training.
Implications and Future Directions
The introduction of neural networks to inertial HPE introduces an avenue to reduce dependencies on high-cost sensor systems and complex setup procedures without diminishing the accuracy required for applications. The low inference times present new opportunities for embedding these models into wearable technology, offering real-time analytics without requiring extensive computational resources.
Future developments should further optimize these neural network architectures, considering larger datasets to enhance generalization and integrate robust sensor-to-segment calibration procedures addressing soft-body and segment offset artifacts more effectively. Further exploration into improved regularization strategies and loss functions specific to rotation matrices may yield additional accuracy enhancements, moving closer to a universal implementation of these technologies in commercial settings.