What it is
AI sports video analysis uses multimodal AI models β models that understand video as well as text β to watch a match the way an analyst would. Instead of a person noting down every goal, foul or break point, the model recognises those moments and returns a structured list: what happened, when, which side it involved, and a short description.
The output is event-based. It answers "what happened, and when?" for every moment that matters in the sport you're analyzing. That makes it the fastest way into your footage: rather than watching everything to find the moments you care about, you start from the list and jump straight to them.
How it works
It starts with the footage: an uploaded file, a direct video URL or a YouTube link. Next comes the sport, which you choose or β for YouTube β auto-detect from the video's title and details. The sport matters more than anything else, because each one is analyzed against its own list of events, its own vocabulary and its own rules about what to ignore. A tennis match and a rugby match have almost nothing in common, and a generic model would treat them the same.
The video is then analyzed by Google's Gemini models, running on Vertex AI. The model returns the events it's confident about β with a timestamp, the team or player, and a description β and leaves out moments it isn't sure of rather than guessing. Finally, the events are laid out beside the video, where you can click any one to jump to it.
Three ways to run an analysis
Standard analysis processes a video in a single pass. It's the right choice for highlights, clips and anything under about 20 minutes.
Parallel analysis splits long footage into short, overlapping segments and analyzes several at once (three by default), adding events to the timeline as each segment finishes. For a full match it's dramatically faster, and if one segment fails the rest still complete.
Stream analysis works on a match while it's being broadcast live. At regular intervals β every 30 seconds by default β it takes the most recent stretch of the broadcast, analyzes it and adds the events to a live feed β and it can overlay those detections on the stream for your viewers.
What gets detected, sport by sport
Football: goals with scorer, assist and finish type; shots; set pieces; cards; defensive actions; offsides; VAR decisions; substitutions.
American football: touchdowns by method, field goals, extra points and two-point conversions, interceptions and fumbles, sacks, significant penalties, punts and returns, and big plays of 20+ yards.
Basketball: field goals by type and free throws, assists, offensive and defensive rebounds, turnovers including shot-clock violations, steals, blocks, fouls and timeouts β with fast-break and clutch moments labelled.
Tennis: aces, double faults and service winners; winners and errors by forehand or backhand; break points created, saved and converted; net play; and Hawk-Eye challenges.
Padel: bandejas, vΓboras and ganchos; smashes por tres and por cuatro; glass play; volleys; lobs; and the golden point.
Pickleball: dinks and speed-ups, volleys and resets, ATPs and Ernes, serves and returns, kitchen violations, and how each rally ended β including third-shot drops versus drives.
Volleyball: serves, attacks and kills, blocks and digs, sets and assists, long rallies, and substitutions and libero rotations, with combination plays identified.
Baseball: hits by type and field location, outs, stolen bases, walks, wild pitches and balks, errors and double plays, and runs scored β with the count and base situation when they're visible.
Cricket: wickets by dismissal type, boundaries and sixes by shot, extras, appeals and DRS reviews, milestones, and dropped catches and misfields.
Ice hockey: goals by strength, penalties by infraction, power plays, key saves, big hits and fights.
Rugby: tries and how they were created, conversions, penalty and drop goals, penalties by offence, cards, and significant turnovers.
Reviewing the results
AI analysis is fast, but you should still check it β and the tools for that matter as much as the detection itself. In PitchView every event has a review status β Pending, Correct, Wrong, Uncertain, Duplicate or Partial β that you set as you check it, with an optional comment. You can edit any event, delete it, or add your own manual events, and filter the list by type, team, status, or AI versus manual. If you want to see exactly how the model reached its answer, the prompt can be previewed before the run and the raw AI response is available after it.
Customising the analysis
The built-in instructions for each sport are a strong default, but every team has its own questions. You can write custom prompts β to focus on a player, a phase of play or a pattern you're coaching β and run or re-run an analysis with them. You can also choose the model: Flash is faster and cheaper, Pro gives higher-quality results, and the temperature and output length can be adjusted.
Keeping and sharing your work
Every analysis is saved to your history with the video details and the settings used, so you can come back to it, keep reviewing or re-analyze it. Results can be exported as an Excel spreadsheet (standard analysis) or JSON (parallel analysis). Your teams β with their short names, countries, colours and player rosters β are kept in one place and reused across analyses, and an organisation can have several users working in the same account.
Event-based analysis vs tracking
It helps to be clear about what kind of analysis this is. Event-based analysis tells you what happened and when: the goal, the break point, the sack. Tracking-based analysis follows every player's position throughout the match to produce heatmaps, distances covered and formation data, and usually needs dedicated multi-camera setups or wearable sensors.
PitchView is event-based. It doesn't produce heatmaps or positional data. In return, it works from a single camera or an ordinary broadcast, across 11 sports, without any special hardware β which is what most coaches, clubs and academies actually have.
Getting reliable results
The biggest factor is the footage. One wide, steady camera that keeps most of the playing area in view gives the model the best chance of seeing each event, and higher resolution helps. Pick the two teams from your Teams list so every event is attributed by name, choose parallel analysis for anything longer than about 20 minutes, and review the results before you rely on them.