How Track & Field AI actually analyzes your video (honest version).

Most AI-powered video analysis apps are black boxes. You upload a clip, something happens, and feedback appears. That’s fine, but for a coaching tool, where you’re supposed to trust the feedback and act on it, you deserve to know what’s actually happening under the hood.

Here’s the honest story behind how Track & Field AI analyzes your track and field video, what it’s good at, and where it has limits.

Step 1: Event-aware frame extraction

When you upload a video, the first thing we do is extract still frames. Most video analysis tools extract frames at a fixed rate, say, one frame every half second. For most track and field events, that’s useless. A pole vault plant lasts 200 milliseconds. If you sample at 2 Hz, you’ll miss the plant entirely, or catch it in a motion-blurred frame that’s unreadable.

We use what we call event-aware extraction profiles. Each event, pole vault, shot put, 100m, etc., has its own profile that specifies where in the video frames matter most.

• Pole vault: 2 fps during the long approach, 8 fps from plant through clearance.
• Sprint starts: 10 fps over the first 15 meters, 4 fps thereafter.
• Throws: 2 fps during the wind-up/entry, 12 fps in the final 0.5 seconds around release.
• Long jump: 4 fps during approach, 10 fps across the last 3 steps, 6 fps in flight.

We cap total frames at 50 per analysis to keep latency reasonable and cost manageable. The result: you’re always looking at a sharp, meaningful frame, not a blurry average.

Step 2: Phase detection

Once we have the frames, we identify which frames correspond to which phase of the event. For pole vault, that means tagging each frame as part of the approach, plant, takeoff, swing, turn-and-push, or clearance. For sprints: block start, drive phase, transition, top-end, finish. Each event has its own phase structure.

Phase detection uses a combination of visual cues and temporal position. Pose estimation (we detect the athlete’s joint positions per frame) helps anchor things like “the takeoff frame is the one where the plant foot leaves the ground.” Temporal position helps when poses are ambiguous.

Step 3: Phase-specific analysis

This is where most of the “intelligence” lives. For each phase, we run event-specific analysis prompts against the frames, looking for the things a coach of that event would look for.

For pole vault plant, we ask: is the top hand fully extended overhead at pole-tip contact? Is the plant leg close to vertical at takeoff? Are the hands aligned with the runway? Each of these is an individual check with its own thresholds.

For sprint drive phase, we ask: is the body angle close to 45 degrees at step 3? Is the first foot strike behind or ahead of the center of mass? Is the arm action symmetrical?

Each phase check generates one observation, one why-it-matters explanation, and one drill recommendation. We tag each with a priority, critical, important, or minor, so you know what to act on first.

Step 4: Cause-and-effect linking

Sometimes a problem you see in one phase actually originated in a different phase. A bad takeoff might come from a late plant. A collapsed landing might come from a rushed penultimate step.

When the AI can trace a cause-and-effect relationship, we tag both frames, the cause frame and the effect frame, so you can see the connection. This is particularly valuable for fixes where treating the symptom won’t actually solve the problem.

What it’s good at

• Phase-level technique feedback. “Your plant leg is tilted 14 degrees under vertical at takeoff” is something we can see and measure reliably.
• Consistency checks. Across multiple attempts, we’re good at telling you “your approach varied by 15cm between attempts” type observations.
• Pattern flagging. “You pop up out of the blocks by step 4” is very reliable.
• Educational value. For athletes, coaches, and parents less familiar with event mechanics, the phase-level explanations are genuinely useful.

What it’s not good at (yet)

Being honest here matters more than marketing. Some limits:

• Absolute measurements. We don’t give you exact release velocities or jump heights, we’d need calibration reference objects in frame for that. We give you qualitative direction (“release angle is flatter than optimal”) more reliably than exact numbers.
• Psychological factors. We can see what your body did. We can’t see why. If your takeoff was hesitant because of fear of the pit, that’s not something AI will catch.
• Context beyond a single rep. We don’t watch your whole season. A coach knows you had a knee tweak last week; we don’t.
• Novel mechanics. If you have a genuinely unusual technique, the AI might misclassify it. Elite outliers tend to break more models than beginners do.

What we do with your video

Your video is used for your analysis and stored locally on your device. We don’t train models on user video. We don’t share analyses. We don’t publish anything. No accounts required to use the app.

If you want to delete an analysis, it’s a single tap from the history screen.

Our take on AI vs. human coaching

AI is great at catching specific, measurable things in specific frames. Human coaches are great at everything else, connecting to you as a person, adjusting training based on life context, delivering tough love at the right moment, and recognizing patterns across a season.

Track & Field AI is built to be the “between sessions” tool, the second set of eyes that catches the details your coach can’t always watch, and gives you something specific to bring back to them. It’s not a replacement for great coaching, and we’re skeptical of any AI product that claims to be.