You hit the follow button, start pedaling down the trail, and the drone peels off to hold you in frame. That’s the promise. Understanding how drone subject tracking works comes down to one question the drone asks 30 times a second: where did that person go, and are they still the person I locked onto? Most consumer follow modes answer it with a camera and a neural net, not a magic GPS tether. Get that distinction and you’ll know exactly when a shot will hold and when it’ll drift off to chase a fencepost.
This is a mechanism explainer, not a buying guide. If you fly a DJI, a Skydio, or an HoverAir, the marketing name changes but the underlying pipeline is close enough that the same failure modes hit all of them.
Camera Follow vs GPS Follow
Two different things get sold under the word “follow,” and they fail for opposite reasons.
Camera follow (DJI’s ActiveTrack, Skydio’s subject modes) uses the drone’s own camera to find and track you visually. Nothing on your body talks to the aircraft. The drone sees a shape, decides it’s you, and keeps the gimbal and flight path pointed at that shape. This is what people mean by subject tracking.
GPS/beacon follow ties the drone to a transmitter you carry — a phone, a watch, or a dedicated remote. The aircraft flies toward that coordinate. No vision required. It’ll follow you through a tunnel or into a dark forest because it’s chasing a signal, not a silhouette.
The tradeoff is blunt. Camera follow frames you beautifully and reacts to how you move, but it loses you the instant it can’t see you. GPS follow never loses the signal but frames you loosely, lags on sharp direction changes, and inherits every meter of consumer GPS drift (roughly 3 to 5 meters under open sky, worse near cliffs and buildings). Serious follow rigs blend both: vision for framing, GPS as a fallback heading when the visual lock breaks. When you notice a drone snap back to you after a bad occlusion, that hand-off is usually what saved the shot.
How Computer Vision Tracks You
Visual subject tracking works by drawing a box around you, describing what’s inside that box, and hunting for the best match in the next frame. The drone builds a compact “appearance model” — your color, contrast, and shape against the background — then predicts where the box should move and re-checks the match every frame. It’s tracking a pattern of pixels, not a legal identity.
Here’s the loop in practice.
- You draw the box. In the app you tap yourself or drag a rectangle. That first bounding box is the drone’s entire idea of “the subject.”
- It builds an appearance model. The tracker samples what’s inside the box — dominant colors, edges, texture — and increasingly runs a small neural network that recognizes person as a category, not just a blob of pixels.
- It predicts, then corrects. A motion model guesses where you’ll be next frame; the vision system confirms or nudges the box. Modern trackers lean on the kind of object-detection research documented in benchmarks like the VOT tracking challenge, which is where a lot of this re-acquisition behavior gets stress-tested.
- It re-acquires after a miss. Lose you behind a tree and a good tracker keeps a memory of your appearance, scanning re-entry points to snap the box back on when you reappear.
That last step separates a $400 drone from a Skydio. Cheaper trackers hold a simple color histogram, so a jogger in a similar jacket steals your box. Skydio publishes how its autonomy stack maps the scene in 3D and models subjects with deep networks, which is why its subject tracking re-acquires through partial occlusion better than most. The company frames these as research-and-engineering claims from its own docs; treat them as vendor specs, not independent test results, and MotorBytes hasn’t flown a unit to confirm them.
DJI’s implementation is worth naming too. According to DJI’s own ActiveTrack documentation, the system offers modes like Trace (follow behind), Profile (follow beside), and Spotlight (lock the gimbal but fly manually) — same tracker underneath, different flight geometry on top.
Why Tracking Loses Lock
Trackers break in predictable ways. Learn the four and you can pre-empt most of them.
Occlusion. You ride behind a tree, a truck, a rock. The box has nothing to match, so it either freezes on the last-known spot or drifts onto whatever it sees next. Thin, repeated occlusion (a row of trees) is nastier than one clean pass behind a building, because the tracker keeps half-losing and half-finding you.
Fast turns and hairpins. The motion model assumes you’ll keep roughly the same speed and heading. Cut a 90-degree turn at speed and the prediction overshoots; the box lands where you would have been. This is the single most common way an MTB or ski shot dies.
Low contrast. A gray wetsuit against gray water. A black jacket in shadow. The appearance model needs your pixels to differ from the background, and when they don’t, the box wanders. Backlight is the classic killer — the sun behind you turns you into a silhouette the tracker can’t describe.
Scale changes and clutter. Fly in close and your box balloons; fly out and you shrink to a few pixels. Add a crowd of similar-looking people and the tracker faces an identity problem it can’t reliably solve. This is why follow modes are far more dependable on an empty trail than at a busy skate park.
Two habits fix most of this. Wear something that contrasts with your background (bright against dark, saturated against gray), and give the drone standoff distance so a single obstacle doesn’t wipe the whole frame.
Obstacle Sensing and Follow Modes
Here’s where two systems that seem unrelated end up fighting each other. Subject tracking wants to keep you centered. Obstacle avoidance wants to keep the aircraft alive. When those goals conflict, avoidance wins — as it should — and your perfect tracking shot gets abandoned mid-flight.
Picture a Trace shot down a wooded singletrack. The tracker has you locked. Then a branch enters the flight path. The obstacle system flags a collision, the drone brakes or climbs, and in doing so it either loses your box behind foliage or falls behind. The follow didn’t fail because tracking was bad. It failed because the two autonomy layers gave contradictory orders and the safety layer had priority.
The interaction runs deeper than braking. On drones with full 3D scene mapping, obstacle data actually helps tracking — knowing where the tree is lets the aircraft plan a path around it while keeping the gimbal on you, instead of just slamming the brakes. That’s the difference between reactive avoidance and predictive navigation, and it’s the whole reason we broke down the sensing hardware separately in our explainer on how drone obstacle avoidance works. If you’re choosing a drone to film moving subjects, the obstacle system matters as much as the tracker — read that piece alongside this one.
One setting people ignore: on several platforms you can dial obstacle sensitivity down for tracking runs in open terrain. It makes the drone commit to the shot instead of flinching at tall grass. Do it only where there’s genuinely nothing to hit.
Getting Reliable Tracking Shots
The tracker holds; the framing still looks amateur if the camera isn’t stable. Tracking decides where the drone points. Stabilization decides whether the footage is watchable, which is a separate mechanical system entirely — we cover it in the next piece on how drone gimbal stabilization works. A locked subject on a jittery gimbal is still an unusable clip.
A few field rules that beat any firmware update:
- Set the box on a clean background. Initialize tracking when you’re clearly separated from clutter, not standing in a crowd. The first appearance model is the one everything else depends on.
- Give it standoff. More distance means one tree occludes less of your path and turns look less abrupt to the motion model.
- Fly Profile or Trace, not Spotlight, for action. Spotlight expects you to pilot; the other two automate the flight path around your movement.
- Re-tap after a bad loss. Don’t fight a drifting box. Pause, re-select, and you reset a clean appearance model instead of chasing a corrupted one.
For a wider look at platforms built to hold a subject at distance, our roundup of long-range photography drones breaks down which airframes keep a stable lock when you’re 300 meters out and moving.
FAQ
Does drone subject tracking use GPS or camera?
Most consumer follow modes track visually with the drone’s camera, not GPS. GPS or beacon follow is a separate mode that chases a signal from a device you carry, so it works without line of sight but frames you loosely.
Why does my drone lose me when I turn?
Its motion model predicts you’ll keep the same heading, so a sharp turn makes the box overshoot. Slow slightly into hairpins or increase standoff distance to give the tracker room to correct.
Can drones track you at night?
Poorly. Visual tracking needs contrast between you and the background, and low light collapses that difference. Beacon or GPS follow is the more reliable option after dark.
Subject tracking is a probability engine running on a camera, not a leash — knowing when it’ll break is what separates a clean follow shot from a drone chasing a fencepost. For more plain-language breakdowns of the tech behind autonomous flight, the Drones & UAV desk at MotorBytes covers the mechanisms without the marketing gloss.
