Research notes
What sports video records, what processing can recover, and what a capture brief needs to specify.
- How well can a head camera recover the wearer's own path?Structure from motion recovered a helmet camera's path on a track cycling ride: 182 of 184 frames registered, with metric scale from a 250 m lap.
- How reliably does a pose model find a rider from a fixed camera?A pose-landmark model found a mountain-bike rider in 299 of 300 frames of a tripod clip, and missed lower-body points where the bike hid the legs.
- How closely can two wearers' cameras be aligned?Two wearers' cameras aligned after capture from a shared exit call: an offset of 8.951 s, within 2 frames (67 ms) at the exit.
- How closely in time must a worn camera and a fixed camera agree?Questions to settle before recording synchronised worn and external sports video.
- What can be reconstructed from one helmet camera on a velodrome?From a 22 s helmet-camera excerpt, 182 of 184 frames registered at 0.985 px mean reprojection error, and 16,130 sparse 3D points are shown.
- What does sport offer embodied AI research?How sports video can support research into perception, movement and robot learning.
- How far can a worn stereo pair measure depth?A helmet stereo pair had a matching spread of 0.35 to 0.70 px (robust 1 sigma) on a path surface. The depth spread that implies from 3 to 16 m.
- How should sport recording volume be specified?Separate camera-hours, session duration and coverage when requesting sports video.
- Which rules limit a worn camera in football, rugby and helmet sports?What the Laws of the Game, World Rugby Regulation 12 and NOCSAE's add-on statement say about worn equipment, quoted at source, and how we read them.