Abstract
Accelerometric gait identification systems should ideally be robust to changes in device orientation from the enrollment phase to the deployment phase. However, traditional Convolutional Neural Networks (CNNs) used in these systems compensate poorly for such distributional shifts. In this paper, we target this problem by introducing an SO(3)-equivariant quaternion convolutional kernel inside the CNN. Our architecture (Quaternion CNN) significantly outperforms a traditional CNN in a multi-user gait classification setting. Additionally, the kernels learned by QCNN can be visualized as basis-independent trajectory fragments in Euclidean space, a novel mode of feature visualization and extraction.
Introduction
Gait identification—recognizing individuals by their unique walking patterns—is a promising passive biometric, typically utilizing smartphone accelerometers. However, a significant hurdle is the "catastrophic drop in accuracy" when a device's orientation changes between training (enrollment) and usage (deployment). Standard Convolutional Neural Networks (CNNs) struggle with these distributional shifts because raw acceleration values are tied to the device's coordinate system. Current solutions often discard 3D spatial data or rely on expensive coordinate transforms. This research introduces a Quaternion Convolutional Neural Network (QCNN) that embeds rotation invariance directly into the architecture.
Methodology: Quaternion Equivariance
The proposed QCNN represents acceleration data using quaternions, which are mathematical constructs commonly used for 3D rotations. The core innovation is the use of $SO(3)$-equivariant kernels. In group theory, $SO(3)$ represents the group of all 3D rotations. An equivariant map ensures that if the input is rotated, the output rotates identically, preserving spatial relationships throughout the network's depth.
The network employs a specific mathematical filter: $$f(\hat{q}_0, \dots, \hat{q}_{2l}) = \sum a_i (\hat{q}_i + b_i)(\hat{q}_l + c_i)\hat{q}_i(\hat{q}_l + c_i)^{-1}$$ where $a, b, c$ are learnable parameters. By maintaining equivariance across layers and only taking the magnitude at the final layer, the QCNN achieves rotation-invariant classification without prematurely discarding directional information.
Experimental Results
The QCNN was evaluated against standard CNNs using two primary datasets: a "Cotemporal" dataset capturing real-world "device-flip" scenarios (opposite orientations) and a larger multi-user dataset with 100 users subjected to random 3D rotations.
Robustness to Orientation Shifts
The performance gap in the "Device-Flip" test was stark. While the standard CNN's accuracy plummeted to 4.4% when tested on a different orientation than it was trained on, the QCNN maintained 46.5% accuracy—a 10-fold improvement. Even with data augmentation (training on rotated data), standard CNNs (29.32%) could not match the native performance of the QCNN (33.41%) in multi-user tests.
Efficiency and Interpretability
The QCNN architecture is highly parameter-efficient, requiring only 59k parameters compared to the 213k used by the standard CNN in cotemporal experiments. Furthermore, the kernels learned by the QCNN can be visualized as 3D trajectory fragments. These represent basis-independent "shapes" of movement, such as specific stride curves, providing a novel way to interpret what the model learns.
Conclusion
The QCNN offers a robust solution for rotation-invariant gait identification, significantly outperforming traditional CNNs in non-ideal deployment scenarios. While absolute accuracy levels indicate room for further refinement for high-security applications, the architecture's efficiency and ability to learn coordinate-independent features mark a significant step forward in mobile biometric sensing.
Summary updated