Abstract:Compared with visible, infrared rays, and ultrasonic positioning technologies, the signal acquisition of portable inertial sensors is not affected by the external environment, such as light and occlusion. Therefore, motion tracking based on inertial signals is a promising complement. However, accurate trajectory reconstruction based on inertial sensors is a great challenge due to intrinsic and measurement errors, especially the drift error that exacerbates the accumulative error in trajectory calculation. To address this challenge, we propose a new trajectory reconstruction method, geometric dynamic segmental reconstruction (GDSR), where we treat the movement trajectory as a combination of basic trajectories. To this end, we design a temporal and spatial interaction segmentation approach to decompose the trajectory into basic segments by combining the dynamic feature of inertial measurement unit (IMU) signals with the spatial morphological feature of motion. Accordingly, we design a geometrical model library with undetermined parameters to match these segments. For precise parameter prediction, we propose an extra-supervised learning method that integrates different prediction tasks into one framework, which can not only expand training samples, but also enable different subtasks to compete with each other, thus improving the parameter prediction accuracy of each subtask, thereby accurately approximating the trajectory segments. To quantify the trajectory reconstruction accuracy, we propose the Fréchet spline sliding error (FSSE) and length error ratio (LER) to evaluate curve similarity. The range of FSSE is [0], [2], where 0 means that the two curves have the same shape. The range of LER is [1, $+infty $ ], where 1 means that the two curves have the same length. We test different IMUs in two experimental scenarios and one public available dataset. In all three tests, the FSSE of the GDSR method is less than 0.09, and the LER is less than 1.31, which is significantly better than all the comparison methods.

Dsrf: A Flexible Descriptor For Effective Rigid Body Motion Trajectory Recognition

On Flexible Trajectory Description for Effective Rigid Body Motion Reproduction and Recognition

DSRF: A Flexible Trajectory Descriptor for Articulated Human Action Recognition

RRV: A Spatiotemporal Descriptor for Rigid Body Motion Recognition

MSM-HOG: A Flexible Trajectory Descriptor for Rigid Body Motion Recognition

RISAS: A Novel Rotation, Illumination, Scale Invariant Appearance and Shape Feature.

Point Context: An Effective Shape Descriptor for RST-Invariant Trajectory Recognition

Describing Local Reference Frames for 3-D Motion Trajectory Recognition

On Multiscale Self-Similarities Description for Effective Three-Dimensional/Six-Dimensional Motion Trajectory Recognition

On Double-Level Kernel Self-Similarities for 3D Motion Trajectory Description and Recognition

Enhancing motion trajectory segmentation of rigid bodies using a novel screw-based trajectory-shape representation

Trajectory Space: A Dual Representation for Nonrigid Structure from Motion

Invariant Descriptors of Motion and Force Trajectories for Interpreting Object Manipulation Tasks in Contact

Motion trajectory recognition using local temporal self-similarities

Arbitrary Spatial Trajectory Reconstruction based on A Single Inertial Sensor

Similarity Evaluation of Rigid Body Motion Based on Multiple Features

Learning Deep Trajectory Descriptor for Action Recognition in Videos Using Deep Neural Networks.

Learning Discriminative Trajectorylet Detector Sets for Accurate Skeleton-Based Action Recognition

A Study of Relative Motion Point Trajectories for Action Recognition

Sketch-based Human Motion Retrieval Via Selected 2D Geometric Posture Descriptor

Degrees of Freedom Matter: Inferring Dynamics from Point Trajectories