Abstract:Motion parameters measurement is essential for understanding animal behavior, exploring the laws of object motion, and studying control methods. Nowadays, advanced computer vision based on machine learning technology supports markerless object tracking in 2D videos. However, due to the fact that all objects move in three-dimensional space, this paper introduces a method of measuring motion parameters using 3D pose estimation. First, an enhanced iterative bundle adjustment algorithm is proposed for multi-camera calibration in a multi-camera vision system by adding two control parameters, which dramatically reduces the reprojection error of multi-camera calibration and lays the foundation for high-precision triangulation. Then, a new spatiotemporal loss function is proposed, which considers the relationship between key points that do not constitute limbs, thereby improving triangulation accuracy. The new multi-camera calibration algorithm is evaluated on ChArUco and 3D pose estimation for metronome, planet pendulum, human hand, Koi, and cheetah. The experimental results show that: (1) the two hyper-parameters in the enhanced iterative bundle adjustment algorithm effectively suppress the influence of noise and play a good role in reducing the reprojection error of multi-camera calibration; (2) the spatiotemporal loss function has a strong constraining ability, the time loss can stabilize high frame rate video triangulation to maintain accuracy, while the space loss can improve the accuracy of triangulation for more complex structures; (3) multi-view data fusion is also conducive to improving the accuracy of triangulation. Moreover, the method was successfully applied to some actual measurement scenes: (1) the accurate measurement of the frequency of a metronome; and (2) the success measurement of the movement of a Koi, which conforms to the basic model of fish swimming. Some dynamic measurement results are displayed at https://github.com/wux024/AdamPose .

3D mouse pose from single-view video and a new dataset

Optical Mouse: 3D Mouse Pose From Single-View Video

A Large-Scale Mouse Pose Dataset for Mouse Pose Estimation

Of Mice and Pose: 2D Mouse Pose Estimation from Unlabelled Data and Synthetic Prior

Motion Parameters Measurement of User-Defined Key Points Using 3D Pose Estimation

Mouse Pose Estimation From Depth Images

Learning 3-D Human Pose Estimation from Catadioptric Videos

Animal3D: A Comprehensive Dataset of 3D Animal Pose and Shape

Marker-Less 3d Human Motion Capture With Monocular Image Sequence And Height-Maps

3D Human Pose Estimation from Multiple Dynamic Views Via Single-view Pretraining with Procrustes Alignment

Benchmarking Monocular 3D Dog Pose Estimation Using In-The-Wild Motion Capture Data

SoloPose: One-Shot Kinematic 3D Human Pose Estimation with Video Data Augmentation

MonkeyPosekit: Automated Markerless 2D Pose Estimation of Monkey

STPoseNet: A real-time spatiotemporal network model for robust mouse pose estimation

Recovering Accurate 3D Human Pose in the Wild Using IMUs and a Moving Camera

Monocular 3D Human Pose Estimation In The Wild Using Improved CNN Supervision

3D Kinematics Estimation from Video with a Biomechanical Model and Synthetic Training Data

BKinD-3D: Self-Supervised 3D Keypoint Discovery from Multi-View Videos

Monocular Real-time Hand Shape and Motion Capture using Multi-modal Data

Towards Generalization of 3D Human Pose Estimation In The Wild

3D Human Pose Estimation from Deep Multi-View 2D Pose