Velocity and stroke rate reconstruction of canoe sprint team boats based on panned and zoomed video recordings
Authors: Julian Ziegler, Daniel Matthes, Finn Gerdts, Patrick Frenzel, Torsten Warnke, Matthias Englert, Tina Koevari, Mirco Fuchs
Organizations: Laboratory for Biosignal Processing, Leipzig University of Applied Sciences, Leipzig, Germany · Research Group Canoeing, Institute for Applied Training Science (IAT), Leipzig, Germany · German Canoe Federation, Duisburg, Germany
Pacing strategies, defined by velocity and stroke rate profiles, are essential for peak performance in canoe sprint. While GPS is the gold standard for analysis, its limited availability necessitates automated video-based solutions. This paper presents an extended framework for reconstructing performance metrics from panned and zoomed video recordings across all sprint disciplines (K1-K4, C1-C2) and distances (200m-500m). Our method utilizes YOLOv8 for buoy and athlete detection, leveraging the known buoy grid to estimate homographies. We generalized the estimation of the boat position by means of learning a boat-specific athlete offset using a U-net based boat tip calibration. Further, we implement a robust tracking scheme using optical flow to adapt to multi-athlete boat types. Finally, we introduce methods to extract stroke rate information from either pose estimations or the athlete bounding boxes themselves. Evaluation against GPS data from elite competitions yields a velocity MAPE of 0.011 [0.008 0.014] (Spearman rho=0.974) and a stroke rate MAPE of 0.009 [0.006 0.013] (Spearman rho = 0.975). The methods provide coaches with highly accurate, automated feedback with minimal manual initialization work required, and without requiring sensors.
Figures & tables
Figure 1: Visualization of the reconstructed scene geometry for a K1 canoe race over the distance of 500 m according to matthes_reconstructing_2025 . Orange lines correspond to lane boundaries, blue ones to boundaries between equidistant segment, and yellowish dots where orange and blue lines cross are expected buoy location alongside the track. See Sec. 3.1 for an outline of the scene reconstruction method. Numbers in red coloured annotations at boat tips correspond to the distance travelled of each boat, which can further be utilized to reconstruct velocity profiles. Best viewed on screen.
Figure 2: Shown is the scene geometry of a regatta course (left) and an approximate orthogonal camera view on the scene (right). For clarity, yellow buoys highlighted in the scene are those visible in the right image. Corresponding buoys in both views can be identified based on their lane boundaries Bi and track segments Sj . Correspondences are used to estimate a homography and propagated to adjacent frames. Adapted from matthes_reconstructing_2025 .
Figure 3: Visualization of all boat classes in our main dataset. C refers to canoe, where single-bladed paddles are used. K refers to kayak, where double-bladed paddles are mandated. The number indicated the number of athletes per boat.
Distance
K1
C1
K2
C2
K4
Total
200 m
8
3
–
–
–
11
500 m
21
6
7
5
4
43
Total
29
9
7
5
4
54
Table 1: Contents of our main dataset.
Figure 4: Graphical Illustration of Automatic Boat Tip Offset Calibration. (1) For every detected athlete bounding box (green), we assume the centre of the bottom edge to be the image position of that athlete (blue point). (2) Athlete positions are transformed to the 2d world space, and a standard offset is applied to the mean position of the athletes (orange). (3) This position is reprojected into the image, and defined as centre of ROI. (4) ROI is input into a trained U-Net, which predicts the exact tip location (bright green). (5) Exact tip location and athlete positions in the image are again transformed to world coordinates. (6) Individual athlete offsets are calculated (red arrows). Best viewed on screen.
Figure 5: Principle of boat tip localization using a U-Net. The image annotation (green dot) of the boat tip (left) is encoded via a heatmap with gaussian kernel (middle). The U-Net is trained to predict this heatmap (right) from which the predicted position (red cross) is derived using argmax over the predicted heatmap.
Figure 6: Visualization of stroke estimation methods across one paddle stroke. Top: A motion signal is built by calculating the Euclidean distance between the detected shoulder joint (orange) and wrist joint (blue) normalized by bounding box width in every frame. Bottom: The mean brightness of a patch 20 % the width and height of the athlete bounding box (purple) in the left bottom corner is used as the motion signal.
Pixel Error (px)
Accuracy (Acc@ k ) [%]
Mean
Median
1 px
3 px
5 px
2.97
1.72
23.5
81.6
90.8
Table 2: Boat tip detection evaluation on test samples.
Figure 7: Perspective variance of the athlete positioning model relative to race distance. While a minor systematic negative shift in mean error occurs as the camera panning angle increases, the mean absolute error remains below 0.4 m. Standard deviation is consistent. The model maintains sufficient accuracy for boat tip estimation. Best viewed on screen.
Velocity Metrics (Median [Q1 Q3])
Method
RMSE (m/s)
RRMSE
MAPE
Spearman ρ
Linear Interpolation
0.444 [0.143 0.742]
0.073 [0.029 0.150]
0.053 [0.025 0.077]
0.915 [0.882 0.944]
Opt. Flow Athlete Ordering
0.433 [0.135 0.593]
0.073 [0.026 0.107]
0.053 [0.020 0.081]
0.933 [0.816 0.945]
Opt. Flow Interp., causal
0.129 [0.063 0.170]
0.022 [0.011 0.036]
0.013 [0.007 0.022]
0.968 [0.839 0.980]
Opt. Flow Interp., non-causal
0.095 [0.061 0.136]
0.016 [0.011 0.023]
0.010 [0.009 0.012]
0.982 [0.963 0.992]
Table 3: Robust multi-athlete tracking ablation. Details see text.
Velocity Metrics (Median [Q1 Q3])
Boat Type
RMSE (m/s)
RRMSE
MAPE
Spearman ρ
K1/C1
0.079 [0.063 0.112]
0.017 [0.014 0.023]
0.011 [0.008 0.013]
0.975 [0.960 0.985]
K2/C2
0.070 [0.048 0.155]
0.015 [0.009 0.033]
0.011 [0.006 0.017]
0.965 [0.924 0.975]
K4
0.095 [0.061 0.136]
0.016 [0.011 0.023]
0.010 [0.009 0.012]
0.982 [0.963 0.992]
All
0.078 [0.061 0.117]
0.017 [0.013 0.024]
0.011 [0.008 0.014]
0.974 [0.949 0.984]
Table 4: Correlation and error measures for velocity profiles. Details see text.
Figure 8: Agreement analysis between the proposed velocity estimates and GPS ground truth. Left: Bland-Altman plot illustrating the median difference and limits of agreement (red, 2.5th and 97.5th percentiles), indicating minimal bias and only few outliers beyond the confidence bounds. Shaded area shows the confidence intervals of the limits of agreement, cmp. Sec. 3.9.7 . Right: Relative velocity error over the full dataset, demonstrating low deviation from GPS measurements across the entire velocity range.
Figure 9: Results of GPS-based (blue) and video-based (orange) velocity profiles visualized over race distance. Left: 500m. Right: 200m. Both figures underpin the high agreement between video- and GPS-based measurements.
Stroke Rate (Median [Q1 Q3])
Method
RMSE (bpm)
RRMSE
MAPE
Spearman ρ
ViTPose
1.753 [0.868 2.921]
0.014 [0.009 0.022]
0.009 [0.006 0.013]
0.975 [0.921 0.990]
BBox
6.467 [4.399 10.031]
0.070 [0.039 0.088]
0.039 [0.019 0.053]
0.807 [0.590 0.914]
Table 5: Performance metrics and computational efficiency for stroke rate estimation.
Figure 10: Agreement analysis between ViTPose stroke rate estimates and Gyrometric ground truth. Left: Bland-Altman plot illustrating the median difference and limits of agreement (2.5th and 97.5th percentiles), indicating minimal bias and only few outliers beyond the confidence bounds. Shaded area shows the confidence intervals of the limits of agreement, cmp. Sec. 3.9.7 . The two distinct clusters are caused by the substantially different stroke rates of canoe and kayak athletes. Right: Relative stroke rate error over the full dataset, demonstrating low deviation from Gyrometric measurements across the entire stroke rat range. The error range slightly varies over distance which can be attributed to occlusions that consistently occurred (cmp. Fig. 11 ) in the stroke rate validation dataset.
Figure 11: Stroke rate analysis was conducted over all boat classes, but data for multi-athlete boats consistently suffered from occlusions due to several poles that become visible while moving the camera to follow the boats. The occlusion occurred at around 120m (left) and 320m (right) race distance. These correspond to the higher standard deviations observed in our predictions at -380m and -180m, see Fig. 10 .
Figure 12: Results of both Pose-based (green) and Bounding Box based (orange) stroke rate estimation, visualized for two races. The left race shows both methods performing well, as they are both able to capture the form of the profile. On the right, both methods show weaknesses and do not correlate exactly with ground truth data. However, the bounding box method delivers much worse results.
Figure 13: Demonstration of the application of the methodological extensions to analyse team boats as presented in this paper to a K4 canoe race over 500m distance. See. Fig. 1 for annotations. Best viewed on screen.
This paper proposes a calibration-free framework for reliably and effectively estimating vehicle speeds from monocular videos, without relying on roadway features, camera calibration, or roadway-feature-based reference objects. The proposed framework estimates vehicle speeds using a 36-keypoint vehicle template and a homography matrix updated at each frame. A YOLO-based keypoint detection module is trained on diverse datasets, and two estimation strategies are compared: keypoint-only tracking and warped optical flow with dense spatial aggregation. Speed is estimated by projecting displacements into metric space using the homography, with validation conducted on over 400 video clips from roadside and overhead datasets, covering speeds from 30 to 100 mph. The method achieves reliable speed estimation on the VS13 and BrnoCompSpeed datasets, with the warped optical flow method delivering MAEs of 15.0% and 9.7%, respectively, and 77.9% and 93.1% of estimates falling within +/-20% error. After applying a 10% trim to remove edge-of-frame outliers, performance improves to MAEs of 11.7% and 7.6%, with within-+/-20% accuracy increasing to 85.3% and 95.4%. This work addresses key limitations of existing vision-based approaches and enables low-cost and efficient speed enforcement using portable devices such as dashcams and smartphones, thereby supporting citizen-based enforcement programs for traffic safety.
Gaofeng Su, Keya Li, Raja Sengupta +1
Department of Civil & Environmental Engineering The University of California, Berkeley · Department of Civil, Architectural and Environmental Engineering The University of Texas, Austin
Velocity-aware pitch control is standard, but under a broadcast viewport half the players are off screen and on-screen velocities come from a drifting calibration. We ask at which layer of broadcast off-ball analysis velocity changes the answer. Inheriting our off-screen imputation protocol (three Metrica matches, 44 m viewport, block-bootstrap CIs), we score four velocity regimes -- none, viewport-legal observed, true-for-visible, true-for-all -- against a velocity-aware ground truth at three layers: imputation, the control surface, and team verdicts. Velocity is nearly useless for imputation (-0.2 pp against a 12--14 pp velocity-free surface MAE), first-order for the surface (-1.5 to -1.8 pp, 11--15% of that MAE), and ten times smaller for verdicts (-0.12 to -0.19 pp). The velocity that matters is the visible channel: perfect occluded-player velocity adds 2--6% of the visible gain, and no last-seen decay policy we tested exceeds that. Omitting velocity blurs the surface (per-frame |e| 2.2--2.6 pp) with small time-averaged bias (per cell <=0.4 pp), whereas imputation error is a structured bias against the defending team's deep zone (5--9 pp). At a fixed velocity window, a noise ladder of eleven jitter settings, including sigma_v-matched pairs, is ordered to first order by one velocity-noise axis sigma_v with break-even ~1 m/s; eleven SoccerNet-GSR clips from one match through our pipeline measure sigma_v=1.65 m/s yet recover 24--36% of the benefit: 43% of the variance is frame-common, which the surface tolerates, and the residual is heavy-tailed and clustered, which Gaussian controls matched on component RMS do not reproduce (+0.03 vs. +0.36). The share of velocity-free error that velocity removes grows with viewport width (7% at 36 m, 21% at 60 m): fix imputation on tight shots, velocity on wide ones. Code and logs are released.
This paper addresses the problem of reconstructing missing or dropped frames in top-down drone video of autonomous surface vehicles performing structured maritime manoeuvres. We propose a pipeline that converts raw GPS telemetry and a single reference frame into a trajectory-guided video sequence using a pre-trained image-to-video diffusion model, requiring no domain-specific fine-tuning. GPS coordinates from onboard telemetry logs are projected into image space via an equirectangular mapping, producing per-vessel motion cues that condition the SG-I2V diffusion model. The generated frames are evaluated against ground-truth video using perceptual, temporal and trajectory-based metrics, and benchmarked against optical flow extrapolation and RIFE interpolation baselines. SG-I2V produces the most naturally appearing frames among all methods (BRISQUE 25.52, closest to ground-truth 23.64), the most realistic motion magnitude (temporal smoothness 1.14 vs. ground truth 1.42), and the strongest GPS trajectory adherence (9.31px vs. 28.70px for ground-truth, the latter reflecting approximate temporal alignment between footage and GPS logs rather than generation error), demonstrating that trajectory-guided diffusion synthesis is a viable approach to maritime video reconstruction under challenging low-texture, small-object conditions.
Stelio Bompai, Ioannis Kontopoulos, Giannis Spiliopoulos +2
Department of Electrical and Computer Engineering, National Technical University of Athens, Greece · Department of Product and Systems Design Engineering, University of the Aegean, Syros, Greece