In this guide
How Thermal Drones See Heat: LiDAR, mmWave Radar and Thermal Imaging for Inspection
A power line can look perfectly healthy from the ground while a loose connector quietly overheats a few metres above your head. The resistance at that joint turns electricity into heat, the heat cooks the metal, and one day the connection fails — sometimes taking the line down with it. Climbing up to check every joint is slow and dangerous, so utilities send a drone instead: a flying sensor platform that reads heat from a distance, holds a steady altitude over the wires, and keeps flying when fog rolls in. This guide explains the three sensors that make that possible — a thermal camera, a single-point LiDAR and a mmWave radar — how each one works, what each one can and cannot see, and what a student team can honestly build at lab scale.
The Matrice 5T launch, and why inspection drones exist
DJI launched the Matrice 5T on 5 November 2026, a drone built specifically for inspection work. Every figure in this paragraph is a reported figure for the Matrice 5T, not a measurement of any student build: a 1280×1024 thermal imager; LiDAR that spots thin power lines up to a reported 32 m; mmWave radar that keeps working in fog, dust, rain and snow; fisheye cameras that hold position down to 0.5 lux; 11× optical and 176× hybrid zoom; laser ranging to a reported 3,000 m; up to 59 minutes of flight; IP55 weather sealing; and a price around €8,600. The launch matters because it shows the direction of the whole inspection-drone field: several different sensors, each covering the others' blind spots, on one airframe.
Inspection is the job that justifies that sensor stack. A transmission line, a solar farm, a wind turbine blade or a factory roof is large, high up, and expensive to shut down — and the faults that matter most, loose joints, cracked cells and failing insulation, announce themselves as heat long before they fail outright. A drone can scan kilometres of line in one flight and bring back a heat map instead of a guess. The matching student build is published as a project topic: Matrice 5T-Inspired Thermal Inspection Drone with LiDAR Mapping and mmWave Radar Obstacle Avoidance. The rest of this guide explains the technology behind that project.
How a thermal camera sees heat
A thermal camera does not see light. It sees infrared radiation — the heat energy every object above absolute zero gives off. The sensor inside is a microbolometer array: a grid of tiny elements, each one warming up slightly as infrared radiation falls on it, and each one's electrical resistance changing with that temperature. The camera reads the whole grid many times a second and turns the resistance pattern into an image, where each pixel is a temperature reading rather than a colour.
Most thermal cameras for inspection work in the long-wave infrared band, roughly 8–14 µm, because that is where objects near room temperature radiate most strongly, and because ordinary air is fairly transparent there. The camera never measures temperature directly; it measures radiated power and converts it using the Stefan–Boltzmann relationship, which says radiated power grows with the fourth power of absolute temperature. That conversion depends on emissivity — how efficiently a surface radiates. A dull, oxidised connector has high emissivity and reads close to its true temperature. A shiny new aluminium conductor has low emissivity, reflects the cold sky, and can look cooler than it really is. Two parts of the same line at the same temperature can show different colours on screen purely because their surfaces differ, which is why a thermal image is a map of apparent temperature, not a thermometer reading.
A loose connector shows up because of simple physics. A loose joint has higher electrical resistance than a tight one, and the power wasted as heat at a joint is I²R — current squared times resistance. Double the resistance and the joint makes twice the heat at the same current, so a bad joint glows as a bright cluster of hot pixels against the cooler line around it. The pattern matters as much as the number: a single hot pixel could be noise or a reflection, but a cluster of neighbouring pixels, all hotter than the conductor on either side, is the signature of a real fault. That is also why resolution matters. The student build uses an MLX90640 thermal camera with a 32×24 grid — 768 temperature readings. The Matrice 5T's reported 1280×1024 imager holds about 1.3 million. The extra pixels do not make the camera more accurate; the MLX90640 is rated at about ±1.5 °C, which is honest accuracy for a lab build. What resolution buys is identification at distance: with 768 pixels, a hot joint 10 m away may be one or two pixels and easy to confuse with noise, while a million-pixel imager still shows the shape of the joint and the wire on either side. A student drone compensates by flying closer and by looking for clusters, not single pixels.
How single-point LiDAR holds altitude — and why thin wires are hard to spot
The student build carries a TF-Luna LiDAR, a single-point time-of-flight sensor. It fires a short pulse of infrared light downward, waits for the reflection, and converts the round-trip time into a distance: distance equals half the speed of light times the time of flight. It does this hundreds of times a second and hands the flight controller a single number — the distance to whatever is directly below. That one number is what makes altitude hold work. A barometer drifts with wind and weather, and GPS altitude is too coarse for low flight, but a LiDAR reading over flat ground is steady to about ±6 cm across its 0.2–8 m range, so the flight controller can hold a fixed height above the line it is inspecting.
The catch is beam geometry. The laser pulse is not an infinitely thin line; it spreads into a small cone as it travels, so the spot it illuminates grows with distance. A power line is thin — a few millimetres across — and if the beam's footprint is wider than the wire, most of the pulse's energy flies past the wire and only a weak reflection comes back. The sensor may report the ground behind the wire instead, or report nothing at all. That is why thin wires are the hardest target for LiDAR, and why the Matrice 5T's reported ability to spot thin power lines up to 32 m is a serious engineering claim: it takes a narrow beam, a sensitive receiver and signal processing that can pick a faint wire return out of the background. The student build does not attempt that. Its LiDAR points down for altitude, not sideways for wire detection, and its 8 m maximum range is a hard limit — beyond it, there is no reading at all.
How FMCW mmWave radar sees through fog
The third sensor works on a completely different principle. The HLK-LD2410 on the student build is a 24 GHz FMCW radar: frequency-modulated continuous wave. Instead of pulses, it transmits a continuous radio wave whose frequency sweeps upward in a repeating chirp. When the wave bounces off an object and returns, its frequency is lower than the wave currently being transmitted, because the transmitted frequency has moved on during the round trip. The difference between the two frequencies is directly proportional to the distance, so the radar measures distance by measuring a frequency shift.
The LD2410 divides its range into distance gates of 0.75 m, out to about 6 m. Each gate reports whether something is present in that slice of space. That is presence detection, not imaging: the radar can tell you that something is 3 m ahead, but it cannot tell you whether it is a wall, a tree branch or a person. It also detects motion, because a moving object shifts the reflected frequency slightly — the Doppler effect — which is why the same sensor family is used in automatic doors and presence detectors.
Radio's advantage is that it passes through what stops light. Fog droplets, dust and rain are much smaller than a 24 GHz radio wave, which is about 12.5 mm long, so the wave passes through them with little scattering. Visible light and the LiDAR's infrared beam, with wavelengths thousands of times shorter, scatter off the same droplets and go blind. That is why the Matrice 5T's reported mmWave radar keeps working in fog, dust, rain and snow while its cameras and LiDAR degrade. The student build gets the same physics at a smaller scale: the LD2410 keeps reporting obstacles in dusty or smoky air where the visible camera feed goes grey. The trade-off is resolution — a 6 m, 0.75 m-gate radar is a proximity sensor, not a mapper — and the radar cannot identify what it sees, only that something is there.
How the three sensors complement each other
No single sensor does the whole job. Each one answers a different question, and each one is blind where the others can see:
| Sensor | Question it answers | What it cannot do |
|---|---|---|
| Thermal camera (MLX90640) | Where is it hotter than it should be? | Cannot measure true temperature without emissivity; cannot prove a fault on its own. |
| Single-point LiDAR (TF-Luna) | How high am I above the ground? | One distance only; cannot map sideways; misses thin wires; blind beyond 8 m. |
| mmWave radar (HLK-LD2410) | Is something close ahead, even in fog? | Presence only; cannot identify objects; 0.75 m gates, 6 m maximum range. |
| Visible camera (ESP32-CAM) | What does it look like? | Needs light; useless in fog, dust or darkness. |
The combination is what makes inspection work. The radar keeps the drone from flying into a tower in fog. The LiDAR holds a steady height over the line. The thermal camera finds the hot joint. The visible camera confirms what the hot spot actually is — a loose connector, a bird's nest, a patch of sun-heated metal. Remove any one of them and the system gets weaker: thermal alone finds heat but cannot tell a fault from a reflection; LiDAR alone flies well but sees nothing wrong; radar alone avoids collisions but inspects nothing.
The student-scale stand-ins
The student build keeps the sensor architecture and replaces every flagship component with an honest lab-scale stand-in. The comparison below puts the two side by side; every Matrice 5T figure is a reported figure, and every student figure is the design target of the published build:
| Item | Matrice 5T (reported) | Student build (design target) |
|---|---|---|
| Thermal imager | 1280×1024 pixels | MLX90640, 32×24 pixels, ±1.5 °C |
| LiDAR | Thin lines spotted to 32 m | TF-Luna, 0.2–8 m, ±6 cm, downward only |
| Radar | Works in fog, dust, rain, snow | HLK-LD2410 24 GHz, 0.75 m gates to 6 m |
| Flight time | Up to 59 min | 12–15 min |
| Range | Laser ranging to 3,000 m | About 100 m Wi-Fi link |
| Weather sealing | IP55 | None — fair weather, visual line of sight only |
| Price | Around €8,600 | Student budget, built on an F450 frame |
The frame is an F450 quadcopter with a Pixhawk 2.4.8 flight controller running ArduPilot, an ESP32-CAM for the visible feed, and an ESP32 ground station that displays the thermal image and the radar gates. Nothing here pretends to match the Matrice 5T. The point of the build is the architecture: three different sensors, each answering its own question, fused on one airframe. The full parts list, wiring and flight setup are in the project topic: Matrice 5T-Inspired Thermal Inspection Drone with LiDAR Mapping and mmWave Radar Obstacle Avoidance.
Spotting a hot joint in code
The thermal camera hands the ESP32 a 32×24 grid of temperatures — 768 numbers, updated several times a second. The simplest useful program is a threshold scan: walk the grid, count the pixels above a temperature that means trouble, and check whether they form a cluster rather than scattered noise. A loose connector on a live line is typically tens of degrees above the conductor around it, so 60 °C is a reasonable first threshold for a lab demonstration on a heated test wire, not a field diagnosis limit.
# Hotspot scan over one MLX90640 frame (32 rows x 24 columns)
HOT_LIMIT = 60.0 # degrees C: flag pixels above this
MIN_CLUSTER = 4 # need at least 4 hot neighbours to count
def scan_frame(temps):
hot = []
for r in range(32):
for c in range(24):
if temps[r][c] > HOT_LIMIT:
hot.append((r, c, temps[r][c]))
clustered = 0
for (r, c, t) in hot:
neighbours = [p for p in hot if abs(p[0] - r) <= 1 and abs(p[1] - c) <= 1]
if len(neighbours) >= MIN_CLUSTER:
clustered += 1
return len(hot), clustered
# hot_count: pixels above the limit; clustered: pixels with hot neighbours.
# A real fault shows clustered pixels; scattered singles are usually noise.
The cluster check is the important part. A single hot pixel can be a reflection, a warm insect, or sensor noise. Four or more hot pixels touching each other, all hotter than the wire on either side, is the shape a real fault makes. In the ground station, the clustered pixels are drawn in red over the visible camera feed, so the operator sees the hot spot in context instead of a bare number.
Frequently asked questions
Why can't radar identify objects?
Because it only measures distance and motion. A radar return tells you that something reflects radio waves at a given range, but a wall, a branch and a person can all return the same strength at the same distance. Cameras identify things by shape and colour; radar has neither, so it is used for presence and collision avoidance, and the camera does the identifying.
Why can't a thermal image alone prove a fault?
Because a hot spot has innocent explanations. Sunlight heats one side of a conductor, a shiny surface reflects a hot sky, and a legitimately loaded joint runs warmer than an idle one. A thermal image shows where to look; proving a fault takes comparison with the same joint under the same load on another day, or a visible-light photo showing discolouration or a loose fitting. Thermal finds candidates, it does not deliver verdicts.
Can a thermal camera see through fog or rain?
Partly. Long-wave infrared passes through light haze better than visible light, but heavy fog and rain scatter infrared too, and water on the lens blocks the view entirely. The sensor that genuinely keeps working in bad weather is the mmWave radar, because its wavelength is far larger than the droplets. That is why inspection drones carry both: the camera for clear days, the radar for the days the camera cannot see.
Why does LiDAR miss thin wires?
Because the laser spot grows with distance and the wire is thinner than the spot. Most of the pulse's energy passes the wire, the return is weak, and the sensor reports the ground or nothing. Spotting a wire at 32 m, as reported for the Matrice 5T, needs a narrow beam and a receiver sensitive enough to catch a faint return — well beyond a single-point hobby sensor, which is why the student build uses its LiDAR for altitude only.
How far can the student drone fly?
About 100 m on its Wi-Fi link, with a design-target flight time of 12–15 minutes, always within visual line of sight and below 120 m. That is enough to inspect a test rig or a short span of line on a campus, and it is the honest scale for a lab build. Long-range inspection is a different machine with a different radio link and a different budget.
Do I need permission to fly an inspection drone?
In India, drone flights are governed by DGCA rules, and flying near power infrastructure, over people, or beyond visual line of sight needs the appropriate permissions. A student build is flown on a campus or test field with the institution's approval, in visual line of sight, below 120 m, and never near live high-voltage lines without the utility's permission. The rules are part of the build, not an afterthought.
Limitations
- All Matrice 5T figures in this guide are reported figures from the launch, not measurements of the student build and not guarantees of what any other drone achieves.
- The student build is a lab-scale prototype: a 32×24 thermal imager, an 8 m LiDAR used for altitude only, a 6 m presence radar, a 12–15 minute design-target flight time and about 100 m of Wi-Fi range. It demonstrates the sensor architecture; it does not inspect real transmission lines.
- A thermal image shows apparent temperature, not true temperature. Emissivity, reflections and sunlight all shift the reading, so no fault is confirmed from a thermal image alone.
- The radar detects presence, not identity, and the LiDAR measures one distance straight down. Neither maps the environment on its own.
- Temperature thresholds in the code example are demonstration values for a heated test wire, not field diagnosis limits for live equipment.
- The build flies in visual line of sight, below 120 m, under DGCA rules and with institutional approval. It is not flown near live high-voltage infrastructure.