INTRODUCTION: Capturing motion data outdoors is difficult, especially for complex sports such as alpine ski racing. Recently, deep learning-based methods 1,2 have attracted broad attention due to their ability to locate human joints (keypoints) in images and videos. While these methods work well for normal skiing, they do not provide accurate data in complex injury situations 3. METHODS: We developed a new method that improves keypoint detection, especially in injury situations. Specifically, we rotate each image incrementally before processing it by a keypoint detector. Thereby, a distribution of keypoints is generated, providing more information about the keypoints` location than the unrotated prediction only. The best keypoints are then selected via shortest path optimization. The performance of two keypoint detection algorithms (AlphaPose and DCPose) and our proposed method in regular skiing (reg), out-of-balance (oob) and fall scenarios is evaluated on the percentage of correct keypoints (PCK) and mean per joint position error (MPJPE) metrics. RESULTS: Our method improves keypoint detection in all categories (reg, oob, fall) and metrics, especially in out-of-balance and fall situations. Improvements were observed independent of the applied detector, outperforming previous methods 3. Table 1: Comparison of DCPose 1 and AlphaPose 2 without postprocessing vs. Zwölfer et al. 3 vs. our method by a) the percentage of correct keypoints (PCK) and b) the mean per joint position error (MPJPE) metric. DISCUSSION/CONCLUSION: Combining our new method with state-of-the-art keypoint detectors, regular skiing and out-of-balance situations can be considered accurate enough for kinematic analysis. For complex fall scenarios, keypoint estimates should be further improved. This also applies for pose estimation algorithms lifting the 2D keypoints to 3D space, which is the next step for more advanced injury analysis. REFERENCES: 1 Liu, Z. et al., 2021. IEEE/CVF; 2 Fang, H.S. et al., 2017. ICCV; 3 Zwölfer, M. et al., 2021. CV4WS
© Copyright 2023 9th International Congress on Science and Skiing, March 18 - 22, 2023, Saalbach-Hinterglemm, Austria. Veröffentlicht von University of Salzburg. Alle Rechte vorbehalten.