Skip to content
SENSORAIA / DOCUMENTATION

Sensoraia Glossary

You do not need to learn all these words before using Sensoraia. Start with the three names below, then come back whenever a guide uses an unfamiliar term. Each definition explains what the word means here, and why it matters when choosing equipment, running a demo or interpreting a result.

The short version: Sensoraia is the project, MORA is the onboard computer, and USHER is the service that coordinates its sensors. The flight controller still flies the aircraft.

Names and responsibilities | Capture and results | Position and flight | Images and timing | AI and models | Installation | Abbreviations A-Z

Your module: AIcam | CHDKbridge | Diplacam | EARcam | LOOMcam | LWIRcam | PANGEA | Spottercam | MIMIC | Camera Lens

A typical journey: the flight controller requests a capture. USHER asks the selected sensor to do the work, then keeps the observation with its available pose and metadata. A module may also return a detection, position or plan. The flight plan decides whether and how to act on it.

Names And Responsibilities

Sensoraia

The camera-compute project as a whole: sensor applications, their supporting software and the tools to try them on a development PC or use them onboard. It is not the name of a particular camera or computer.

MORA

The role of the aircraft's onboard companion computer, not a specific board. Sensoraia's initial resource-constrained reference is a Raspberry Pi Zero 2 W; another suitable compute board can serve as MORA when its OS, architecture, drivers, interfaces and resources are supported or adapted. It runs sensor and processing software separately from the flight controller. MORA is hardware, not the name of a service to restart.

USHER

The service that coordinates capture and sensor work on MORA, exchanges requests and results with the flight controller, and records information alongside observations. Restarting USHER restarts this software, not the flight controller. You can also run USHER locally on the development PC.

Companion computer

A separate computer that helps the flight controller with jobs such as image processing and sound analysis. In this project, that computer is MORA. "Companion" describes its role, not an alternative autopilot.

FC: Flight Controller; Autopilot

The aircraft's flight-control computer. It estimates aircraft state and runs the control and navigation software. Sensoraia can supply observations or targets; the FC and its flight plan remain responsible for flying the aircraft.

GCS: Ground Control Station

The operator's ground-side application for monitoring the aircraft and sending supported commands. Paparazzi's GCS can display telemetry, photographs' event markers and waypoints. A point shown on its map is not automatically a command to fly there.

Module

A part of the system with a particular job, such as EARcam's sound search or LWIRcam's thermal capture. A Sensoraia module belongs to the sensing software; a Paparazzi FC module belongs to the aircraft software. They are not interchangeable installation targets.

Paparazzi

The open-source autopilot system used by the supplied flight examples. It includes aircraft software, simulation and ground tools. Sensoraia connects to it at the communication boundary; Sensoraia is a separate project.

Sensor

Equipment that measures something about the world. Here that includes visible and thermal cameras and microphones. Other sensor types are possible extension ideas, not a promise that a driver for every device is already included.

SODA

Sensoraia's saved-image processing application. USHER can hand a completed image to SODA for its selected camera-specific processing, review products or optional lens emulation. SODA is not the camera driver, and it does not mean every image is automatically analysed by a neural network.

Capture And Results

API: Application Programming Interface

A defined way for software components to request work or exchange information. An API description explains how to interact with a component; it is not the same as the end-user command or the physical cable carrying a message.

Acknowledgement; ACK And NACK

A reply about a request. ACK is an acknowledgement; NACK is a negative acknowledgement. Always check what was acknowledged: accepting a request, saving a file and confirming a waypoint position are different events. An acknowledgement of admission is not proof of completed sensing.

Backend

The implementation selected to do a job. A camera backend handles capture from a particular source; an inference backend selects how a model runs. For example, choosing an ONNX inference backend does not itself choose which physical camera supplies the image.

Camera ID And Camera Mask

A camera ID identifies one sensor slot. A mask selects one or more slots by setting bits in a number. The protocol has eight addressable slots, but not every slot has an implemented backend. For example, ID 5 is selected by mask 0x10 (decimal 16), not mask 5. Selecting several cameras does not prove their exposures happen at the same instant.

Capture; Shot; Trigger

A trigger requests an observation. A capture is the sensor work that obtains it. Shot is the camera-oriented name used by the shared interface, including for some non-image measurements. A request can be rejected or fail; counting triggers is therefore not the same as counting saved photographs.

DC_SHOT

A Paparazzi camera-event message used for shot numbering and GCS display. One admitted trigger has one GCS shot number even when it selects several cameras. It is not the UART protocol's message-type number, nor independent proof that every selected camera saved an image.

Completion; Saved; Recorded

Completion means a particular stage finished. A saved result concerns an image output; recorded can acknowledge a non-image observation's admission. Neither word alone guarantees a valid geographic target or a finished map. Check the result of the stage you actually need.

Freshness; Stale Result

How recently the information underlying a result was obtained. A newly arrived reply can already be stale if it was computed from an old pose or frame. Freshness limits prevent old information being mistaken for a current target; they do not improve the accuracy of the original measurement.

IPC: Inter-Process Communication; Socket

IPC is how separate running programs exchange information. A socket is one communication endpoint. For example, Diplacam's local worker socket is between software processes on one computer; it is not the FC UART wire protocol or an instruction to change a flight-plan block.

Protocol; Message Type; Payload

A protocol defines how a conversation is encoded and interpreted. A message type identifies the kind of message; its payload carries the associated data. A message-type number is not a camera ID, shot number or waypoint ID, even if two of those numbers happen to be equal.

Session; Sortie

A sortie is one aircraft outing. A software session is a separately identified period of work, such as one search or planning run. A session token helps reject delayed replies from earlier work. A session need not cover an entire flight, and restarting one need not restart the aircraft.

Telemetry

Information reported by the aircraft or another running component, such as position, status or the waypoint coordinates the FC currently holds. It tells you what was reported; receiving telemetry is different from sending a command.

Timeout; Backpressure

A timeout ends a wait that has taken too long; backpressure means a component cannot currently accept more work. A busy camera can reject a request instead of adding it to an ever-growing queue. Retrying is a separate policy: a timeout does not prove a physical exposure never happened.

UART: Universal Asynchronous Receiver/Transmitter

The serial communication interface used between the FC and MORA. TX sends, RX receives and GND provides the common electrical reference. For the documented 3.3 V connection, TX connects to the other device's RX. A compatible connector shape does not establish compatible voltage levels.

Valid; Invalid; Unavailable

Valid means the stated checks for a particular result passed. Invalid means it cannot be used as that result; unavailable means the required information was not supplied. An unavailable temperature is not zero degrees, and a valid computation is not an accuracy certificate.

Worker; Asynchronous Processing

A component that carries out a job separately from the request-handling loop. Asynchronous means the requester need not wait in place until all work is done; the result arrives later. It does not mean the result is immediate, nor that unlimited jobs can be queued.

Position And Flight

AGL: Above Ground Level

Height above the terrain underneath the aircraft, not height above sea level. An aircraft 50 m above a hill is at 50 m AGL even when the hill itself is high above the sea. AGL matters when estimating a camera footprint or a sound source's distance.

Attitude: Roll, Pitch And Yaw

The aircraft's orientation. Roll is banking, pitch is nose-up or nose-down tilt, and yaw is rotation around the vertical axis. Camera projection needs orientation as well as position: a banked camera generally sees different ground from a level camera at the same coordinates.

Coordinates; Local Origin

Numbers that locate a point within a declared reference system. Latitude and longitude use angles; a local east/north map uses distances from a chosen origin. An east offset of 20 m has no geographic meaning until you know that origin. Keep both the reference system and the units with the numbers.

ENU And NED

Two ways to name local coordinate directions: East, North, Up and North, East, Down. Their component order and vertical sign differ. In NED, a positive down value points downward, not upward. Neither is the aircraft's forward/right body frame.

Fillet turn

A rounded corner joining two straight flight legs. Rather than stopping or turning sharply at their meeting point, the aircraft follows a connecting circular arc. This changes which ground positions a sound-search pass samples.

Flight Plan; Block; Stage

The mission instructions interpreted by the autopilot. In Paparazzi, a block groups a mission activity and a stage is a step within it. Updating a target waypoint's coordinates does not, by itself, select a new block or restart the current stage.

Geolocation; Geotag

Geolocation estimates a geographic position, for example where an image pixel meets the ground. A geotag stores geographic information with an image. The aircraft's position in EXIF is not automatically the position of every object in that photograph.

GNSS And GPS

Global Navigation Satellite System is the general name for satellite navigation systems. Global Positioning System is one such system; "GPS" is also commonly used informally for the aircraft's satellite-derived position. Coordinate precision in a file does not establish position accuracy.

Heading And Course

Heading describes where the aircraft's nose points. Course describes the direction of travel over the ground. Wind can make them different, so course is not a substitute for camera yaw when projecting an image.

HOME; HOLD

Named waypoints in the example flight plans. HOME is the plan's home reference; HOLD is a configured waiting location where provided. They are not necessarily the aircraft's current position. Entering a HOLD block does not automatically mean hovering exactly where a failure happened.

MSL: Mean Sea Level; Ellipsoid Height; Geoid

Different references for altitude. MSL is a sea-level-based reference; ellipsoid height uses a mathematical Earth surface. A geoid model relates these references. For a simple example, ground at 30 m MSL and an aircraft at 80 m MSL give 50 m AGL. Do not subtract an MSL ground height from an ellipsoid aircraft height without the appropriate conversion.

NPS: New Paparazzi Simulator

Paparazzi's flight-simulation environment. It runs the aircraft software with a simulated aircraft so you can exercise a complete command-and-result loop without a physical flight. A successful simulation is not a measurement of real sensor accuracy or onboard processing speed.

Pose

Position and orientation together. A useful sensor pose also says when it applies and which coordinate conventions it uses. The pose at request time may differ from the pose during exposure because the aircraft keeps moving.

UTM: Universal Transverse Mercator

A family of projected coordinate systems that express position in metres within numbered zones. UTM is convenient for local distance calculations, but an easting/northing pair must retain its zone and hemisphere.

Waypoint; Target

A named geographic point available to the flight plan. A sensing target may be stored as a waypoint, but detecting something and commanding the aircraft to fly toward it are separate actions. The flight plan chooses the navigation behaviour and flight altitude.

WGS84: World Geodetic System 1984

The Earth reference used for the latitude/longitude coordinates in these workflows. It provides a common geographic reference, not a guarantee that a reported position is correct to any particular number of metres.

Units: m, mm, Hz, fps, ms, us And ns

Metres and millimetres measure distance; 1 m is 1,000 mm. Hertz, Hz, counts cycles or updates per second; fps counts frames per second. Milliseconds, microseconds and nanoseconds are fractions of a second: 1 s is 1,000 ms, 1,000,000 us or 1,000,000,000 ns. A camera's advertised frame rate is not a promise of the saved-photo or inference rate.

Images And Timing

Calibration And Validation

Calibration estimates parameters or corrections using reference measurements. Validation checks performance on independent observations. A calibration file existing, or a fitting program finishing, is not the same as demonstrating accuracy on new measurements.

EXIF: Exchangeable Image File Format

Structured metadata that can be stored inside a photograph, including supported location, camera and exposure information. Sensoraia also preserves specific module-owned evidence records. Editing software may discard metadata; keep original files when you need their measurement history. More about Sensoraia's EXIF.

Exposure; Shutter Time; Rolling Shutter

An exposure is the interval over which a sensor gathers light. Shutter time is its duration. With a rolling shutter, different image rows are exposed at different times. A moving aircraft can therefore introduce blur or geometric skew even when one pose is stored for the whole image.

Footprint; FOV: Field Of View

FOV is the angular extent of a camera's view; the footprint is the area of ground that view covers at a particular position, height and orientation. A wider lens or greater AGL generally covers more ground. Horizontal, vertical and diagonal FOV are different measurements, not interchangeable versions of one number.

Frame; Sequence Number

One image from a camera stream. A sequence number identifies its place in that stream and helps associate results with the correct observation. Camera sequence numbers and FC shot numbers are different identifiers; matching numeric values alone does not prove two records belong together.

Ground Plane; Nadir

A ground plane is a flat surface used to simplify image projection. Nadir means directly downward. A flat-ground calculation cannot account for every roof, tree or slope; a downward-looking camera can still move away from nadir when the aircraft banks.

JPEG, PNG, RGB And Grayscale

JPEG is a commonly compressed photographic image format; PNG is a lossless image format often used for diagrams and masks. RGB means red, green and blue colour channels; grayscale stores intensity without colour. These describe image storage or appearance, not whether the pixels are calibrated temperatures. JPEG expands to Joint Photographic Experts Group; PNG to Portable Network Graphics.

Latency; Jitter

Latency is delay, for example between requesting a photo and its exposure. Jitter is variation in that delay. File arrival time includes later processing and transfer, so it is not automatically the exposure time.

Metadata; Provenance

Metadata describes an observation: for example its pose, time, sensor or model identity. Provenance records where it came from and how it was processed. These help distinguish physical observations, simulated inputs and derived results without guessing from a filename.

Observation; Request Pose; Observation Pose

An observation is what a sensor measured, with its relevant context. Request pose describes the aircraft when the capture was requested; observation pose describes the estimated aircraft pose at sensing time. The latter requires timing evidence. Missing timing must not be silently relabelled as a synchronized observation.

Pixel; Resolution

A pixel is one sample in an image grid. Image resolution commonly means its width and height in pixels; it is not the same as the detail the optics resolve or the accuracy of a geographic estimate. Enlarging an image creates more output pixels, not new measured detail.

Sidecar; Embedded Data

A sidecar is a separate file associated with an image, such as an analysis INI file. Embedded data lives inside the image file itself. Keep required sidecars with their original image; do not assume every workflow writes them or that an image editor preserves every embedded record.

Synchronization; Clock Alignment

Relating events or clocks so the system can estimate which aircraft pose belongs to a measurement. This includes limits on timing uncertainty, not just putting timestamps on both files. At 10 m/s, a 0.1 s timing error alone corresponds to about 1 m of travel.

Timestamp; UTC; Monotonic Time

A timestamp says when something happened according to a particular clock. UTC, Coordinated Universal Time, is a shared civil-time reference. Monotonic time measures elapsed time without ordinary wall-clock jumps. Monotonic values from different computers cannot simply be subtracted as though their clock origins were the same.

Uncertainty, Precision And Accuracy

Uncertainty describes how well a value is known. Precision concerns repeatability or, in file output, the number of digits shown. Accuracy is closeness to an independent reference. More decimal places, a sharper image or a high model score do not by themselves establish better location accuracy.

AI And Models

AI, ML And Neural Network

Artificial Intelligence is the broad field of software performing tasks such as recognition. Machine Learning learns behaviour from examples. A neural network, or NN, is one kind of learned model. In Sensoraia, a network's job depends on its contract: detecting objects, classifying sound or proposing measurements are different tasks.

Class; Classification

A class is a label such as "vehicle" or an acoustic category. Classification chooses or scores categories for an input. It does not necessarily say where an object is in an image, and it does not automatically produce a geographic target.

Confidence; Score; Threshold

A model's score expresses support for an output according to that model. A threshold decides which scores to accept. Do not read 0.8 as an 80% chance of a correct result unless the model's calibration establishes that interpretation. Computational-validity flags and geographic accuracy are different quantities again.

CPU And Neural Accelerator

The Central Processing Unit runs general-purpose programs. A neural accelerator is specialized hardware for compatible model computations. Sensoraia's ONNX baseline runs on the CPU; an IMX500 RPK uses the AI Camera's neural processor. A model file alone does not make an accelerator available.

Dataset; Ground Truth

A dataset is a collection of examples. Ground truth is the reference answer associated with an example, such as a labelled object or a known synthetic field value. Training and evaluation need suitable labels, rights and representative inputs; saved imagery is not automatically a ready-to-use training dataset.

False Positive And False Negative

A false positive reports something that is not there; a false negative misses something that is. Adjusting a score threshold changes this trade-off. One successful demonstration does not measure how often either occurs in different light, terrain, sound or weather.

Inference And Training

Training learns model parameters from examples. Inference runs the already trained model on a new input. Selecting an ONNX model and analysing a photo performs inference; it does not train or automatically improve that model.

Model; Graph; Checkpoint

A learned model combines computations with trained parameters. An exported graph describes those computations in a runtime-readable form. A checkpoint stores a particular training state. Models are not interchangeable just because their filenames have the same extension.

Model Descriptor; Model Contract

The descriptor states which graph and tensor layout an application expects, including preprocessing and a graph hash. The wider contract says what the inputs and outputs mean. A vehicle detector is not a visual-return model, and a generic feature vector is not a navigation direction. Choose a compatible model bundle.

ONNX And ONNX Runtime; ORT

Open Neural Network Exchange is the model format. ONNX Runtime, often abbreviated ORT, is the engine that executes compatible graphs. Sensoraia applications share native runtime support but select their own models and settings. ONNX support does not imply support for every custom operator.

Preprocessing; Normalization; Letterbox

Preparing input exactly as a model expects. Normalization rescales values; letterbox resizes an image while preserving its aspect ratio and pads the remaining area. Colour order and interpolation also matter: a plausible-looking image is not enough to establish the right numerical input.

FP32, INT8 And Quantization

FP32 uses 32-bit floating-point numbers. INT8 uses 8-bit signed integers. Quantization represents values at reduced numerical precision, usually with scales and offsets. It may reduce model cost, but speed and accuracy depend on the export and runtime. An INT8 filename alone is not proof of compatibility, acceleration or unchanged quality.

RPK

The model-package file used by the IMX500 camera path. It must suit the camera and application contract; it is not a general-purpose CPU model. An .onnx and an .rpk are different deployment choices, not extensions to rename.

Tensor; Shape; NCHW And NHWC

A tensor is a structured array of numbers passed to or from a model. Shape specifies its dimensions. For image batches, NCHW means batch, channels, height, width; NHWC means batch, height, width, channels. The same numbers arranged differently can mean a completely different input.

Installation And Operation

ABI: Application Binary Interface

The low-level rules that let a compiled program use libraries and operating system facilities. Two ARM64 files can still be incompatible if their runtime requirements differ. Matching the processor name alone is not a deployment compatibility check.

Architecture: x86-64, ARM64 And AArch64

The processor instruction family a binary runs on. ARM64 and AArch64 refer to the 64-bit Arm target used by MORA; x86-64 is the development-PC target in the supplied builds. The production source is shared, but the compiled binaries are different.

Build; Cross-compile; Deploy

To build is to turn source code into runnable binaries. To cross-compile is to build for another target, such as compiling ARM64 on the PC. To deploy is to upload and activate a prepared release on MORA. Building or packaging does not deploy it, and deployment can restart USHER. Follow the installation path.

CLI; Option; Flag; Environment Variable

Command-Line Interface means operating a program through typed commands. An option or flag, such as --help, changes that invocation. An environment variable, such as AIRCRAFT, supplies context inherited by programs. Neither automatically rewrites an application's settings file.

INI, JSON, CSV And XML

Text formats for different kinds of information. INI is an initialization- file format used for settings; JSON, JavaScript Object Notation, stores structured values; CSV, Comma-Separated Values, stores rows of data; XML, Extensible Markup Language, uses named elements, including Paparazzi flight plans. A JSON file does not require a JavaScript runtime.

Firmware

Software running inside a device such as a camera or flight controller. Uploading a Sensoraia application to MORA does not replace Canon firmware, install CHDK on its SD card or flash a new autopilot image.

Headless; SSH

Headless operation uses no locally attached display or keyboard. Secure Shell, SSH, provides a remote terminal. A command typed in an SSH session runs on the remote computer; a command typed in your ordinary PC terminal runs on the PC. This distinction matters for build and deployment commands.

Log; Diagnostic; Exit Status

A log records what software reports while running. A diagnostic explains a condition or failure. Exit status is the final numeric result of a command, normally zero for success. Output text alone may describe an intermediate stage; inspect the saved result and final status too.

Native

Compiled machine-code execution, here primarily C and C++. In a build command, "native build" also means a build for the current PC rather than a cross-build. Native does not mean real-camera input, hardware acceleration, no dependencies or demonstrated whole-board memory headroom.

Package; Release

A prepared collection of application binaries and required settings, models, runtime libraries and notices. A release identifies a particular collection for distribution or deployment. A package on the PC is not already activated on MORA.

Absolute Path; Relative Path; Working Directory

An absolute path starts at the filesystem root, such as /home/air/sensoraia/bin/earcam. A relative path needs a reference directory. Shared model-settings paths are relative to their INI file; ordinary command arguments may instead be relative to the shell's working directory. Moving a settings file can therefore change what its paths select.

RAM; CMA; MiB

Random Access Memory is working memory shared by the operating system, applications and device buffers. The Pi Zero 2 W reference board has 512 MB total; other MORA boards may have different capacities. Linux's Contiguous Memory Allocator, CMA, reserves contiguous memory for devices such as cameras. MiB, mebibytes, counts bytes in powers of two. A small executable file does not imply a small runtime memory requirement.

SDK; Toolchain; Sysroot

A Software Development Kit supplies build-time components for a target. A toolchain includes its compiler and related tools; a sysroot provides target headers and libraries. Sensoraia prepares these on the development PC for cross-compilation. They are not the application runtime package deployed to MORA.

Service; systemd; Journal

A service runs under operating-system supervision rather than only inside an interactive terminal. On MORA, systemd manages usher.service; its journal contains service logs. Editing a service configuration is different from rebuilding the application or changing a camera's firmware.

SHA-256; Hash; Checksum

A hash is a fingerprint computed from file contents; SHA-256 is Secure Hash Algorithm 256-bit. A checksum comparison can reveal changed or mismatched files. It does not establish trustworthy authorship by itself: an attacker could replace both an untrusted file and its untrusted expected hash.

USB, UVC, PTP And CSI

Universal Serial Bus, USB, connects devices such as microphones and thermal cameras. USB Video Class, UVC, and Picture Transfer Protocol, PTP, are different device protocols carried over USB. The Camera Serial Interface, CSI, is the ribbon-cable camera connection used by the AI Camera. These are not substitutes for the FC-to-MORA UART link.

AIcam

AIcam is the visible-camera path using the Raspberry Pi AI Camera. Spottercam and LOOMcam can use the same physical camera for different processing jobs. AI Camera setup and operation.

IMX500

Sony's image sensor with an integrated neural processor, used in the Raspberry Pi AI Camera. Its hardware can run a compatible RPK while capturing images. Ordinary image capture does not automatically enable a particular detector.

libcamera; rpicam-still

libcamera is the camera-control framework used by the native camera workers. rpicam-still is a separate Raspberry Pi camera application used by the stock capture path. They are not two names for the same executable. A physical camera still needs one capture owner at a time.

Sensor Mode; Crop; Binning

A sensor mode chooses how the camera reads its image. A crop selects part of the sensor; binning combines samples to produce a lower-resolution readout. Changing these can change image geometry, so a calibration for one mode must not be assumed valid for another merely by matching the filename.

CHDKbridge

CHDKbridge controls supported Canon cameras from MORA or the PC. Canon-camera setup and commands.

CHDK: Canon Hack Development Kit

A separately installed software add-on running inside a supported Canon camera. It adds scripting and control functions. CHDK must match the camera model and firmware; installing Sensoraia on MORA does not install CHDK on the camera.

CHDKbridge

Sensoraia's native controller that communicates with camera-side CHDK over USB/PTP. It runs on MORA or the PC, not inside the Canon camera. USHER's CHDK backend invokes it for the selected camera work.

Direct Download

The optional CHDK capture path that transfers JPEG data directly over USB when the camera port supports it. Ordinary capture saves the photo on the camera's SD card and downloads that file. CHDK support alone does not prove the direct-JPEG capability is available.

Optical Zoom; Normalized Zoom

Optical zoom changes the lens's magnification. CHDKbridge's normalized 0..255 control maps onto the camera's available zoom steps; 128 is roughly the midpoint of that control range, not a guaranteed 2x magnification. This is different from cropping or enlarging pixels digitally.

Playback Mode; Lens Retraction

Playback is the camera mode used for reviewing images rather than taking them. Requesting it can let the powered Canon camera retract its lens. Mode confirmation does not measure physical lens position or retraction time; a later capture can extend the lens again.

Diplacam

Diplacam proposes where to measure next using accumulated knowledge of a field, rather than following only a fixed coverage pattern. The measurement-to-plan guide.

Belief; Mean; Variance

The planner's current estimate of the field and its uncertainty. The mean is the estimated value at each grid cell; variance describes uncertainty in that estimate. Belief is not the complete field truth: it changes as measurements arrive.

CMA-ES: Covariance Matrix Adaptation Evolution Strategy

A numerical search method that improves candidate plans by evaluating and adapting a population of proposals. Diplacam offers it as an explicitly selected comparison mode, not a silent substitute for a failed ONNX model. It is unrelated to Linux camera-memory CMA.

DDIM: Denoising Diffusion Implicit Model

A sampling method that repeatedly refines a noisy candidate using a diffusion model. Diplacam runs twenty native denoising steps. One ONNX call predicts noise for one step; it is not the entire eight-point planner by itself.

Diffusion Planner

A learned planner that turns an initial noisy candidate into proposed measurement locations, conditioned on the current belief and aircraft position. "Diffusion" here describes the model's generation process, not physical spreading of a measured substance.

GP: Gaussian Process; Posterior

A statistical model that relates nearby field values and maintains an estimate with uncertainty. After incorporating a measurement, the updated estimate is the posterior. Measuring one area can change expectations nearby, but does not turn unmeasured cells into independently observed truth.

Informative Path Planning

Choosing measurement locations to learn useful information efficiently, rather than merely visiting evenly spaced positions. The research objective combines expected information and travel cost. Whether a particular model achieves useful plans needs evaluation in the intended conditions.

P1-P8; Plan Generation; Snapshot

The eight waypoints refreshed by the Diplacam example. A generation identifies one completed native plan; a snapshot keeps its eight points together during retrieval. Publishing eight ordinary GCS waypoint moves is not an atomic eight-point update of the FC.

Scalar Field; Synthetic Observation

A map with one numerical value at each location. Diplacam's supplied CSV stands in for sensor returns: only the current footprint is sampled into the GP belief. It is not an RGB image, a full spectral-image cube or an already implemented driver for a physical sensor.

Training Range; Predicted Z

The conditions represented during model training. Diplacam's bundled model uses XYZ internally and was trained for 10..40 m heights; the current demo commands 2 m and publishes only XY. Predicted Z is discarded as a flight command. Correct software execution at 2 m does not establish planning quality outside the model's training range.

EARcam

EARcam searches by sound: it associates microphone measurements with aircraft positions to estimate where an alarm source is located. Continue with the sound-search guide.

ALSA: Advanced Linux Sound Architecture

The Linux sound system that supplies microphone samples to EARcam. An ALSA device name selects an audio input; it does not select an acoustic detection algorithm.

Band-pass filter

A filter that keeps a chosen range of sound frequencies and reduces sounds outside it. EARcam uses this to concentrate on the alarm's frequency range rather than treating all microphone noise as equally useful.

Contrast: EARcam tone contrast

A comparison between the strongest detected tone and the energy across the selected frequency band. A distinct alarm tone has stronger contrast than broad noise. This is an audio quality check, not visual image contrast.

Decibels: dB, dBFS And dB SPL

A logarithmic way to compare signal levels. The reference matters: dBFS means decibels relative to a digital recording's maximum level; dB SPL means sound-pressure level relative to a standard physical pressure. A microphone's dBFS reading is not automatically a calibrated dB SPL reading. Keep the microphone gain consistent when comparing EARcam measurements.

Envelope

A smoothed description of how loud a signal is over time. It bridges short gaps between an alarm's pulses so one brief quiet interval does not make the source appear to vanish.

Gauss-Newton

A repeated-improvement method for fitting a mathematical model to measurements. EARcam uses a fitted sound model to estimate a source position from readings taken at different aircraft positions. It does not simply choose the coordinates of the single loudest reading.

Goertzel filter

An efficient way to measure how much of a particular tone is present in a sound recording. EARcam examines several nearby frequencies to accommodate variation in the alarm tone.

Hampel filter And MAD: Median Absolute Deviation

A method for spotting measurements that differ unusually strongly from nearby ones. MAD measures typical spread around the median, so one extreme spike does not set the scale for everything else. EARcam uses this kind of robust filtering to reduce the effect of isolated bad readings.

Huber weights

A way to reduce the influence of measurements that disagree strongly with a model, without giving up on the whole collection. In EARcam's position fit, this helps prevent a few unusual sound readings from dominating the answer.

motionSCOUT

MSA's personal motion-alarm device, used as the sound source in the EARcam search example. Its pulsed alarm gives the detector a recognizable tone pattern to look for. It is the target sound source, not a component running on MORA.

Sound picture

The JPEG heat map produced after an EARcam search. It makes the sound-search result easy to inspect and associate with flight metadata. It is a computed visualization, not a photograph taken by a microphone.

PCM: Pulse-Code Modulation; Sample Rate

A representation of recorded sound as regularly spaced numerical samples. The sample rate counts samples per second, not how often a position estimate is produced. An audio model must receive the declared rate and window length; changing only a file's label does not resample its audio.

RMS: Root Mean Square

A measure of signal level across a window of samples. EARcam's raw RMS level and an optional classifier's probability describe different things. A loud sound need not be the desired alarm, and a high classifier score is not a calibrated sound-pressure measurement.

Star Refinement

An acoustic-search pattern with passes in several directions around a candidate location. The different viewing positions help refine the sound source fit; it is not a new type of microphone. The flight plan still owns the actual path and its constraints.

LOOMcam

LOOMcam is the experimental visual-return application. Its model integration and bounded target policy are present; a trained homing network is not included. Visual-return settings and workflow.

Features; Embedding

Numbers produced by a network to describe an input. An embedding arranges descriptions so they can be compared in the intended feature space. Features from different models may not be comparable. A feature vector alone does not specify a return direction or a geographic position.

Homing; Visual Return

Using visual information to help identify a direction toward a known home region.

Reference Photographs; Outbound Phase

Images retained during the outward part of the example flight for possible later visual comparison. References carry their available capture and model evidence. Keeping these photos is not proof that a model can reliably match them from a different viewpoint.

Return Vector; Vector Frame

A declared model output describing a direction toward the return spot. Its frame states whether components mean Earth north_east, aircraft body_forward_right (rotated into Earth coordinates using the photo's EXIF IMU roll/pitch and magnetometer heading), or simulated (an explicitly labelled homeward vector seeded by the photo's neural features for testing before a model is trained). LOOMcam checks and converts the declared frame; it never guesses the meaning from two raw numbers.

T1; Look-ahead; HOME Protection Radius

LOOMcam's example return waypoint, initially at HOME. Look-ahead places a target along an accepted homing ray; the protection radius (60 m by default) locks T1 at HOME once the aircraft arrives so it can settle into a smooth 80 m standby orbit without chasing a point underneath its belly. These checks bound target updates; they do not demonstrate visual homing accuracy by an untrained model.

LWIRcam

LWIRcam captures supported thermal cameras. Measurement and geolocation depend on the camera's available data and the enabled analysis features. Capture guide | Thermal calibration guide

LWIR: Long-Wave Infrared

The infrared wavelength region used by these thermal cameras. A displayed thermal image represents detected infrared information; its colour palette is not visible-light colour, and brightness is not automatically a calibrated temperature reading.

Radiometric; Temperature Plane; Image-only

Radiometric data supports numerical temperature-related measurements. A temperature plane contains those per-pixel values separately from the display image. Supported Tiny1C capture preserves this plane; the supported TDK image-only mode does not provide it. A colourful thermal JPEG alone cannot supply missing temperature samples.

Kelvin64

The preserved Tiny1C temperature encoding: kelvin multiplied by 64. For a valid sample, divide by 64 to obtain kelvin, then subtract 273.15 for Celsius. Reserved invalid codes are not cold or hot temperatures and must not be converted into ordinary measurements.

Emissivity; Reflected Temperature

Emissivity describes how effectively a surface emits thermal radiation compared with an ideal emitter. Reflected temperature accounts for thermal radiation reflected from the surroundings. A shiny surface can show reflected heat rather than its own temperature; visual brightness cannot resolve that ambiguity on its own.

FFC And NUC

Flat-Field Correction and Non-Uniformity Correction address uneven response across a thermal sensor. A camera may use a shutter/reference event as part of its correction. This is different from lens calibration and does not by itself establish absolute-temperature accuracy. Do not assume a software readback records every correction event.

Intrinsics; Extrinsics; Boresight

Intrinsics describe the lens and sensor geometry, including focal lengths, image centre and distortion. Extrinsics describe how the camera is mounted relative to the aircraft. Boresight concerns the viewing-axis alignment. Both lens geometry and mounting matter when placing a hotspot on a map.

Palette; Display Window; Percentiles

A palette converts displayed levels into colours. A display window chooses the temperature range mapped into those levels. Percentiles can choose a range from the distribution of valid values. These can make a scene easier to inspect without changing the preserved temperature samples.

ROI: Region Of Interest

A selected part of an image or sensor plane. LWIRcam can measure a rectangular ROI, point or line using native sensor coordinates. Rotating the displayed JPEG does not rotate those measurement coordinates; check the coordinate convention before selecting a region.

Hotspot; Ground Projection

A region meeting the thermal detector's selection conditions. Ground projection estimates where its image position intersects the assumed ground using calibration and observation pose. It is an estimated location, not proof that the region is a fire or that a surveyed error requirement has been met.

Correction Profile; Hold-out Check

A correction profile records measured corrections for a specified thermal device and configuration. A hold-out check uses reference measurements not used to fit those corrections. Lens, mounting and temperature checks solve different problems; one successful calibration does not replace the others.

PANGEA

PANGEA, Photo-Assisted Neighboring Geographic Edge Assembly, turns saved photographs and their supported positioning information into photo maps. It is a separate processing job, not a prerequisite for every capture. Photo-map workflows.

Georeferenced

Associated with a geographic coordinate system so image positions can be placed on a map. A georeferenced output can still have errors from pose, timing, lens assumptions or terrain; the presence of coordinates is not a survey-accuracy certificate.

GSD: Ground Sampling Distance

The ground distance represented by an image pixel, often expressed in centimetres per pixel. A map at 10 cm/pixel does not automatically have 10 cm positioning accuracy or recover detail finer than its source images.

Mosaic; Orthomap

A mosaic combines several images into one visual map. An orthomap places imagery in a map-like, top-down geometry. PANGEA's reference-plane workflows do not reconstruct every terrain height, so elevated objects can still be displaced or show seams.

Seam; Blending; Exposure Compensation

A seam decides where overlapping source images contribute. Blending smooths their visual transition; exposure compensation reduces brightness differences. These improve appearance but do not correct all geometric misalignment or establish absolute colour/temperature calibration.

SIMPLE And GREAT

Names of PANGEA's two completed-survey modes. SIMPLE uses a simpler compositing approach; GREAT adds graph-cut seam selection and multiband blending. They describe processing choices, not formal accuracy grades. The separate live photo-map workflow has its own input and resource rules.

Coverage; Gap Filling

Coverage says which ground was represented by accepted photographs. Gap filling makes output visually complete using supported source-colour completion. Filled pixels are not extra photographs or newly measured terrain. An output with no blank pixels need not have complete observed coverage.

Incremental Map; Provisional Revision; Finalize

An incremental map updates as accepted photos arrive. A provisional revision is a coherent intermediate product, not a complete survey. Finalizing ends input collection and produces the finished session output; it does not independently validate geographic accuracy.

Manifest; Checkpoint; Resume

A manifest identifies inputs or packaged files. A checkpoint records completed work and the identities needed to reuse it. Resume continues a matching interrupted job rather than starting from zero. Editing committed inputs can invalidate that match; a checkpoint is not permission to mix unrelated surveys.

XYZ Tiles; Tile Pyramid; Zoom Level

A web-map layout addressing square image tiles by zoom level and X/Y indices. A tile pyramid stores several zoom levels for efficient viewing. Higher zoom shows a finer grid; it does not create finer source detail. XYZ here means tile addresses, not a three-dimensional point or elevation model.

MIMIC

MIMIC is Map-tile Imagerie Modeling in Interactive Camera-sim: the map-based camera simulator used in the PC examples. Make camera views from cached imagery.

Map Cache; Map Tile

A map tile is one small image from a map provider. The cache is the local collection already downloaded by the GCS. MIMIC uses the configured cached aerial imagery to render a camera view; displaying the aircraft in the GCS does not automatically prove every tile in its camera footprint is available.

Mock; Fake Camera; Simulated Detection

A mock replaces a particular component for a test. A fake camera can supply saved or rendered images while the real native model still runs. A simulated detection instead fabricates a detector result under an explicit demo policy. These are different substitutions; a mock JPEG does not mean every later processing stage is fake.

Rendered Camera View

An image generated from the simulated camera's pose and ground imagery. Turning or banking the aircraft changes what it sees. MIMIC's ground texture is not a physical exposure, and visible aerial imagery rendered for a thermal path is not a temperature measurement.

MORA-in-the-loop; Camera-in-the-loop

In MORA-in-the-loop, a simulated flight exchanges requests with USHER on the real board. Camera-in-the-loop includes supported physical capture hardware. These test different boundaries; real capture paired with simulated pose does not establish real-flight geolocation accuracy.

Random Seed

A starting value that makes a pseudo-random demo repeatable under the same conditions. It helps compare software behaviour across runs. Seeded random targets are still invented targets, not observations made by a trained detector.

Camera Lens

The optional lens emulator adds declared optical and shutter effects to supported MIMIC imagery while retaining the clean source. Lens profiles and examples.

Lens Distortion; Fisheye

Lens geometry that bends the relation between a scene direction and an image position. A fisheye deliberately uses a wide-angle projection with strong distortion. Correct projection needs that lens's model, not just a wider crop of a pinhole image.

Pinhole Model; Ray

A simplified geometric camera model. A ray is an imagined line from the camera through an image pixel; intersecting it with the assumed ground locates that pixel's footprint. The model describes geometry, not measured focus, lens blur or actual shutter timing.

Lens Profile; Derivative Image

A lens profile declares the emulation's optical and exposure parameters. A derivative is the processed image made from an original. Labelling an emulated photo preserves its meaning; it does not turn assumed parameters into a calibrated physical camera.

Spottercam

Spottercam detects objects, with the supplied application focused on vehicles. Its image detections and geographic target updates are separate steps. Vehicle-search guide.

Bounding Box; Detection

A bounding box encloses a detected object in image coordinates. Detection reports a class, box and score. It does not by itself locate that object on Earth; geographic projection also needs the corresponding camera geometry, pose and ground assumptions.

IoU And NMS

Intersection over Union measures overlap between two boxes: overlap area divided by their combined area. Non-Maximum Suppression reduces competing boxes for the same apparent object. These operate on image detections, not on the aircraft's navigation waypoints.

T1-T5; Best-Five Targets

Five target waypoints in the Spotter demo, initially at HOME. The native store fills empty slots, then replaces the lowest score only with a strictly higher one. These are not LOOMcam's single return target T1, and moving them does not automatically interrupt the survey.

Zero Detections; Analysis Failure

Zero detections means a model ran successfully without accepted objects. Analysis failure means the result could not be produced or trusted. They are not interchangeable. A genuine empty result must not trigger fabricated detections; the explicit random spotting policy belongs only to the simulation harness's documented fallback cases.

Abbreviations A-Z

Each link leads to the explanation and its practical meaning above. Names such as MORA, USHER and SODA are project names; their roles are more useful than an invented expansion of their letters.

Term Meaning Or Place To Start
ABI Application Binary Interface
ACK / NACK Acknowledgement / negative acknowledgement
AGL Above Ground Level
AI / ML / NN Artificial Intelligence / Machine Learning / neural network
ALSA Advanced Linux Sound Architecture
API Application Programming Interface
ARM64 / AArch64 The 64-bit Arm build target
CHDK Canon Hack Development Kit
CLI Command-Line Interface
CMA Linux Contiguous Memory Allocator
CMA-ES Covariance Matrix Adaptation Evolution Strategy
CPU Central Processing Unit
CSI Camera Serial Interface
CSV Comma-Separated Values
dB / dBFS / dB SPL Decibels; the digital or physical reference matters
DDIM Denoising Diffusion Implicit Model
ENU / NED East/North/Up; North/East/Down
EXIF Exchangeable Image File Format
FC Flight Controller
FFC / NUC Flat-Field Correction / Non-Uniformity Correction
FOV Field Of View
FP32 / INT8 32-bit floating-point / 8-bit signed integer
fps Frames per second
GCS Ground Control Station
GNSS / GPS Global Navigation Satellite System / Global Positioning System
GP Gaussian Process
GSD Ground Sampling Distance
Hz Hertz; cycles or updates per second
INI / JSON Settings format / JavaScript Object Notation
IoU / NMS Intersection over Union / Non-Maximum Suppression
IPC Inter-Process Communication
JPEG / PNG / RGB Image formats and red/green/blue colour channels
LWIR Long-Wave Infrared
MAD Median Absolute Deviation
MiB / RAM Mebibytes / Random Access Memory
MIMIC Map-tile camera simulation
ms / us / ns Milliseconds / microseconds / nanoseconds
MSL Mean Sea Level
NCHW / NHWC Image-tensor dimension orders
NPS New Paparazzi Simulator
ONNX / ORT Open Neural Network Exchange / ONNX Runtime
PANGEA Photo-Assisted Neighboring Geographic Edge Assembly
PCM / RMS Pulse-Code Modulation; Root Mean Square
PTP / UVC Picture Transfer Protocol / USB Video Class
ROI Region Of Interest
RPK The IMX500 model package
SDK Software Development Kit
SHA-256 Secure Hash Algorithm 256-bit
SSH Secure Shell
TX / RX / GND Transmit / receive / ground
UART Universal Asynchronous Receiver/Transmitter
USB Universal Serial Bus
UTC Coordinated Universal Time
UTM Universal Transverse Mercator
WGS84 World Geodetic System 1984
XML Extensible Markup Language
XYZ Map-tile addresses, not height data