Sixth Computer Vision Winter Workshop Bled Slovenija February
Sixth Computer Vision Winter Workshop, Bled, Slovenija, February 7 -9 2001. Errors and Mistakes in Automated Player Tracking Janez Perš 1 , Marta Bon 2, Stanislav Kovačič1 1 Faculty of Electrical Engineering, University of Ljubljana, Slovenia 1 Faculty of Sport, University of Ljubljana, Slovenia janez. pers@fe. uni-lj. si, marta. bon@sp. uni-lj. si, stanek@fe. uni-lj. si http: //vision. fe. uni-lj. si/
Outline • Introduction and motivation • Sources and types of errors • Tracking parameters • Ground truth, experiments and results • Conclusion
Introduction and motivation Automated player tracking system has been developed in our lab. How accurate is it?
Introduction and motivation Related work: 10 articles describing people/sport related tracking at ICPR 2000, Barcelona. • 8 did not mention tracking accuracy • 1 proposed the accuracy measure • 1 specified the accuracy (tracking the tennis ball) Research is focused mainly on reliability. Our motivation: sport experts (end users).
Sources and types of errors Note: • We track global movement of players for the purpose of match analysis. • No operator mistakes are allowed.
Tracking parameters • Tracking method used: • 3 available • 2 included in experiments • Post-processing of trajectories • filter width • Position of the players: • radial distortion • non-uniform quantization • Activity of the players: • movement of extremities
Ground truth • A reliable reference for player position was needed. • Several patterns were drawn on the court, players were instructed to follow them. • Patterns were measured using measuring tape.
Experiment I. Position accuracy (RMS error) and path length error with respect to: • different tracking methods • player position • player activity. • 2 players at court boundary, 3 near court center. • 60 seconds (@25 fps) standing still, 150 seconds active (passing ball, jumping - all at the same position).
Experiment I. • 2 D histograms of player position, cumulative histograms of absolute position error: RMS position error: 0. 18 - 0. 50 m (still) 0. 28 - 0. 63 m (active) Path length error: 0. 6 - 1 m/player/min (still) 6 - 10 m/player/min (active) • Combined (color + template tracking) method performed better than background subtraction. • Wide FIR filter decreases error in path length - heavy smoothing preferred.
Experiment II. • Effect of FIR filter width (wide filter = intense smoothing) to square trajectory. • Error measure: • In presence of rapid changes in player direction, wide filter distorts trajectory.
Experiment III. • Velocity accuracy (RMS error) for uniform player motion. • Circular trajectory as ground truth, constant velocity of players assumed. • Reference velocity calculated from trajectory length and the time player needed for one round. • 5 players, Vref from 2. 7 to 3. 2 m/s • RMS error of player velocity: less smoothing: 0. 21 to 0. 35 m/s (6%-12%) more smoothing: 0. 07 to 0. 20 m/s (3%-7%)
Experiment IV. • Velocity and position accuracy. • APAS (manual video-based kinematic analysis tool) was used as a ground truth. • Two short sequences only - motion analysis using APAS is time consuming. RMS position error: RMS velocity error: 0. 28 m to 0. 38 m 0. 5 m/s to 0. 7 m/s
Conclusion • RMS position error: from 0. 2 m to 0. 6 m • RMS velocity error: from 0. 2 m/s to 0. 4 m/s • Path length error: from +1 m to +10 m (per one player per minute) • Players are large non-rigid objects - at this scale the limit is imposed by the definition of player position, velocity and path length itself. • Ideas for future work: above definitions and similar research concerning manual position measurements. |
- Slides: 13