Glossary

Every technical word used in this section, in one sentence. Nothing here is specific to our project unless it says so.

The model

Object detector

A program that looks at a picture and draws a rectangle around each thing it recognises. Ours recognises exactly one kind of thing: a landing pad.

Model

The trained program itself: a file full of numbers that the computer uses to make its decision. "Training a model" means adjusting those numbers until the answers get good.

YOLO / YOLO11n

A well-known family of object detectors. The n stands for nano, the smallest and fastest version. We use the smallest one because it has to run on a tiny computer.

Class

A category the model can recognise. Ours has one class, called landingPad. A model with two classes could recognise pads and people.

Bounding box

The rectangle the model draws around what it found — four numbers: left, top, right, bottom.

Inference

Running the finished model on a new picture to get an answer. Training is the slow part done once; inference is the fast part done thousands of times.

Input size

The picture size the model expects, ours 320 × 320 pixels. Camera pictures are shrunk to this before the model sees them. Feeding a different size makes the model worse, even though it still "works".

The data

Dataset

The collection of pictures used to teach the model, each one with the correct answer marked by hand.

Label / annotation

The hand-drawn marking that says "the pad is here in this picture". Ours are polygons (outlines), not just rectangles.

Train / validation / test split

The dataset is cut into three parts. The model learns from the training part, is checked during learning on the validation part, and is finally graded on the test part, which it must never have seen. Otherwise the grade is meaningless.

Generalise

Whether a model works on pictures it has never seen before, rather than just on the ones it learned from. A model can be good in two very different ways: it memorised the training pictures, or it learned what a landing pad actually looks like. Only the second is useful, because in flight every picture is new. Everything on the Evaluation page is an attempt to tell the two apart.

Leakage

When pictures from the test part are also, effectively, in the training part — for example two near-identical photos taken a second apart. The model then scores well by remembering, not by understanding. This happened to us; see Dataset.

Augmentation

Generating additional training pictures by altering existing ones: flipping, rotating, blurring, darkening. It teaches the model that a pad is still a pad when it looks slightly different.

Hard negative

A training picture that contains no pad at all. Without these, the model never learns that "there is nothing here" is a valid answer. Adding them was the single most useful thing we did.

Roboflow

A website for storing, labelling and exporting datasets. Ours lives there.

Measuring it

Confidence

How sure the model is about one detection, from 0 to 1. 0.78 means fairly sure.

Threshold

The confidence below which a detection is discarded. Set too low it admits noise; set too high it misses real pads.

True positive / false positive

A true positive is a real pad the model found. A false positive is the model claiming a pad where there is none — it once reported a rucksack as a landing pad.

Recall

Of all the pads that were really there, what fraction did the model find? 1.00 means it found every one.

Precision

Of everything the model claimed was a pad, what fraction really was one?

IoU (Intersection over Union)

How well the model's rectangle overlaps the correct one, from 0 (no overlap) to 1 (perfect). Usually a detection counts as correct at IoU ≥ 0.5.

mAP

The standard school grade for an object detector: one number between 0 and 1, where 1 is perfect. It rolls two things into one — did the model find the objects that were there, and did it avoid inventing ones that were not. When someone says "the model scores 0.99", they mean mAP. : It comes in two strengths: : - mAP50 — was it found? A detection counts as correct if the rectangle overlaps the real pad by at least half. - mAP50-95 — was it found, and tightly enclosed? Considerably stricter, so this figure is always lower. Our model scores 0.995 on the first and 0.81 on the second: it always finds the pad, but the rectangle is sometimes loose. : In our project this grade turned out to be useless for choosing between models, because all six attempts scored almost exactly the same. See Evaluation.

FP32 / INT8

How precisely the model's numbers are stored. FP32 is full precision — the original model as trained on a laptop. INT8 stores the same numbers much more coarsely, which makes the model roughly four times smaller and much faster, at a small cost in accuracy. The drone needs INT8; the laptop does not.

Epoch

One complete pass through all the training pictures. "80 epochs" means the model looked at every picture 80 times.

Checkpoint / weights

The saved file produced by training: the model itself. Ultralytics calls the best one best.pt, which is why a folder can end up with several different files all called "best".

Saturated metric

A score that no longer discriminates because every candidate reaches close to the maximum. Ours saturated at 0.995.

Probe / stress test

Our own replacement for the saturated score: deliberately spoil the test pictures the way flying spoils them — rotate them, shrink them, blur them, and see which model survives.

Nadir

Looking straight down. The drone's camera points nadir; all our training photos were taken at an angle, by hand.

Getting it onto the drone

Raspberry Pi Zero 2 W

The small computer on the drone. It talks to the flight controller and runs the mission logic. It is too slow to also run an AI model.

IMX500 / Raspberry Pi AI Camera

A camera with a small AI processor built into the sensor itself. The model runs inside the camera, and the Pi only receives the finished answer. This is the whole reason we can do AI on this drone at all.

NPU

Neural Processing Unit: the AI processor inside the camera sensor.

Quantisation / INT8

Shrinking the model by storing its numbers with less precision (whole numbers instead of decimals). It makes the model much smaller and faster, at a small cost in accuracy. Required for the camera chip.

.tflite

TensorFlow Lite: the model format for running on a normal processor, such as the Pi's own CPU. Our fallback, not our main route.

.rpk

Sony's package format. The only format the IMX500 camera chip can load. A .tflite file cannot be used here.

packerOut.zip

The half-finished result on the way from a trained model to an .rpk.

ONNX

A general model-exchange format. We use it to simulate the quantised model on a laptop, which is how most of the "quantised" numbers in Evaluation were measured: a good stand-in, but not the real sensor.

NMS (Non-Maximum Suppression)

Cleaning up the model's raw output, which usually contains several overlapping rectangles around the same object, down to one. On the IMX500 this already happens inside the camera.

Letterboxing

Shrinking a picture to the model's input size by adding grey bars instead of squashing it, so nothing gets distorted. Getting this wrong silently ruins the results.

imx500-package

The tool that builds the final .rpk. It only runs on ARM Linux: a Raspberry Pi or an ARM build server, never a Mac or a normal PC.

Flight side

Companion computer

The Raspberry Pi, as opposed to the flight controller. It makes decisions; the flight controller flies.

Flight controller

The board that actually controls the motors and keeps the drone stable.

MAVLink

The language the Pi and the flight controller speak to each other.

EKF

The flight controller's internal estimate of where the drone is and how high. It combines several sensors. Our detector depends on its height estimate to convert "pad is 30 % to the right" into "pad is 1.2 m to the right".

GUIDED mode

The flight mode in which the Pi, not the pilot, tells the drone where to go.

Milestone 2

Our staged bring-up step in which the drone hovers and only logs what the camera sees, without acting on it. The safe place to check that the detector's numbers are right.