Investigating Multi-view Human Pose Estimation for Shooting Posture Analysis : With Evaluation against Inertial Sensor Ground-Truth Data

Sammanfattning

Inertial motion sensors (IMU) are commonly employed in posture assessment tasks. However, is it possible to capture the same information via other medium? The aim of this thesis was to explore how computer vision-based pose estimation compares against pose estimation utilizing IMUs for the use of standing shooting posture assessment. An additional aim was to investigate which of the chosen camera placement configurations score the best according to the following performance metrics: MPJPE, PCP3D, AP_0.1, AP_0.15, and AP_0.2. Out of these, MPJPE was chosen as the main evaluation metric as it is the error of the predicted keypoint location from the ground-truth. To guide this exploratory study, two research questions were formulated, namely: (RQ1) How do computer vision–based human pose estimation and motion sensor–based data compare in terms of accuracy for shooting posture analysis? and (RQ2) Which of the selected camera configurations, defined by unit count and placement, achieves the lowest Mean Per Joint Position Error for shooting posture estimation? To answer the two research questions, 20 participants, 10 of which had received formal training and 10 without, participated in an experimental data collection sequence. The experimental data collection sequence consisted of a shooting sequence in a simulated shooting environment using a weapon replica whilst wearing a motion capture suit. The participants were recorded by six cameras in two different predefined layouts. A state-of-the-art pose estimation model was subsequently trained using the data gathered from the cameras with the data from the 17 IMUs in the motion capture suit as the ground-truth to compare the predictions of the model against. To properly explore the impact of camera placement, multiple different configurations were investigated. Out of these, the best performing camera layout scored results outside the defined acceptable error range. This discrepancy could be attributed to factors such as inconsistent camera setups, environmental factors such as lighting, or poor camera calibration. Although the performance fell short of the acceptable range, a noticable difference between performances in camera configurations could still be observed. This study contributes by showing that state-of-the-art models in human pose estimation perform well on controlled benchmarking datasets but are still difficult to adopt for real-world applications and proving that camera position has an impact on pose estimation performance which is rarely evaluated in pose estimation research.