Technology · AI tracking

AI Camera Tracking: Onboard Detection and Target Following

Detection, lock and follow running on the payload rather than in the ground station

Tracking is the difference between a camera and a sensor. A payload that can lock an object and hold it in frame frees the operator to watch the object instead of the aircraft, and it frees the autopilot from a command loop that would otherwise fight every gust.

Person / vehicle / vessel / aircraft
20+ simultaneous targets
Anti-occlusion logic
Payload-side compute

Request a quotation All product categories

An ai tracking camera does two jobs in sequence: it decides what in the frame is worth following, then it keeps the line of sight on that object while the airframe moves. Both happen on the payload, using the visible channel, the thermal channel or both, so the behaviour an operator sees does not depend on link quality or on a laptop in a ground station.

A drone tracking camera that only follows motion loses the object the moment it stops, and a drone tracker camera that only follows a manually drawn box makes the operator hold a target through rotor wash at 30x zoom. A working drone camera tracking loop combines detection with a motion model, re-acquires after occlusion, and reports the target position back to the ground station as a coordinate.

How the loop is built

Detection classes
Person, vehicle, vessel and aircraft detection trained on the bands the payload actually uses, so night-time tracking runs on thermal rather than on a low-light visible frame.
Lock and follow
The operator gates a target with a box or a tap; the tracker then keeps the line of sight closed on the object while the gimbal absorbs airframe motion.
Motion model and re-acquisition
A motion model predicts where the target will be when it passes behind a tree, a mast or another vehicle, which is what stops a track breaking on occlusion.
Multi-target tracking
Twenty or more objects can be tracked simultaneously, with the displayed lock switching between them, so a single aircraft can watch a group rather than an individual.
Georeferenced lock
With the laser rangefinder active, the tracked pixel becomes a range and a coordinate, which the ground station can pass to a map, another aircraft or a response team.
Link-loss behaviour
Detection and tracking continue if the data link drops, and the payload holds the target until the link returns, instead of drifting off the object while nobody is watching.

Tracking modes and when to use them

The same payload runs several modes; the mission decides which one is armed.

Mode	What the payload does	Use it when
Manual gate	The operator draws a box on any object; the payload follows it.	The target is not in the detection classes, or a specific object must be followed.
Click to track	A single tap locks the nearest detected object.	Fast-moving surveillance where the operator reacts rather than selects.
Auto detect	The payload finds and reports all targets in the field of view.	Wide-area search before a specific object has been chosen.
Multi-target	Tracks a group and reports each object separately.	Crowds, convoys, traffic and vessel groups.
Scan pattern	The gimbal runs a search pattern while detection runs continuously.	Route and perimeter sweeps where nothing specific is expected yet.
Geo lock	Tracks an object at a coordinate rather than a screen position.	Handing a target to a second aircraft or to a ground team.

Four payloads with the full tracking module

The tracking module is the same across the range, so an operator who has flown one payload can fly them all.

AI tracking camera with onboard detection and target following

HQ · 3-axis · EO + IR + LRF

HQ-135T

The flagship of the compact range: 40x zoom, thermal imaging and 3 km ranging on one stabilised cage.

Visible: 1920 × 1080, 40x zoom 4.25 – 170 mm
Thermal: 640 × 512, 35 mm fixed
Laser: 15 m – 3 km (± 1 m)
Weight: ≤1.4 kg

HQ · 3-axis · EO + IR + LRF

HQ-125TB

Balanced mid-range turret: 30x zoom, a long thermal lens and 3 km ranging.

Visible: 1920 × 1080, 30x zoom 4.3 – 129 mm
Thermal: 640 × 512, 35 mm fixed
Laser: 15 m – 3 km (± 1 m)
Weight: ≤1.2 kg

XQ · 2-axis · EO + IR + LRF

XQ-200T

A continuous-zoom thermal channel from 40 mm to 220 mm plus 9.5 km ranging on a 5 kg pod.

Visible: 1920 × 1080, 30x zoom 4.3 – 129 mm
Thermal: 640 × 512, continuous zoom 40 – 220 mm
Laser: 50 m – 9.5 km (± 1 m)
Weight: ≤5 kg

SQ · 2-axis · EO + IR + LRF + designator

SQ-200T

Laser designation plus 8 km ranging on a 5.5 kg two-axis pod.

Visible: 1920 × 1080, 30x zoom 4.3 – 129 mm
Thermal: 640 × 512, continuous zoom 40 – 220 mm
Laser: 300 m – 8 km (± 1 m)
Weight: ≤5.5 kg

Compare the range

Model	Series	Axes	Configuration	Visible camera	Thermal imager	Laser rangefinder	Weight
HQ-135T	HQ series	3-axis	EO + IR + LRF	1920 × 1080, 40x zoom 4.25 – 170 mm	640 × 512, 35 mm fixed	15 m – 3 km (± 1 m)	≤1.4 kg
HQ-125TB	HQ series	3-axis	EO + IR + LRF	1920 × 1080, 30x zoom 4.3 – 129 mm	640 × 512, 35 mm fixed	15 m – 3 km (± 1 m)	≤1.2 kg
XQ-200T	XQ series	2-axis	EO + IR + LRF	1920 × 1080, 30x zoom 4.3 – 129 mm	640 × 512, continuous zoom 40 – 220 mm	50 m – 9.5 km (± 1 m)	≤5 kg
SQ-200T	SQ series	2-axis	EO + IR + LRF + designator	1920 × 1080, 30x zoom 4.3 – 129 mm	640 × 512, continuous zoom 40 – 220 mm	300 m – 8 km (± 1 m)	≤5.5 kg

Why the compute sits on the payload

The link is the weakest part of the system

A ground-station tracker depends on a video downlink and a command uplink. Both can be degraded by terrain, distance or interference, and the failure is silent: the operator sees a frozen or degraded image and the track is already lost. Running detection and tracking on the payload removes the round trip from the control loop.

Latency and control stability

Every stage of latency in a tracking loop makes the gimbal work harder and the image less stable.
Payload-side ai camera tracking keeps the loop at sensor frame rate rather than at link frame rate.
Commanding the payload to follow a target is one small message; streaming video to a ground station to compute the same result is not.

What the operator still controls

Detection proposes, the operator decides. Target classes, minimum object size, exclusion zones and whether tracking may re-acquire automatically are all set at the ground station, so the payload never escalates a track on its own.

Tracking works with the rest of the payload, not instead of it.

Long range imagingHolding a lock at full zoom.Read more →
Night visionDetection on the thermal channel after dark.Read more →
EO/IR pod systemsTurrets with the full tracking module and long rangefinders.Read more →
Laser designator podsTracking plus 1064 nm designation.Read more →
Evaluating and integrating an AI payloadWhich functions to specify, how to test detection and tracking separately, and what to request from a supplier before you order.Read the buyer’s guide →
Procurement and supplier evaluationHow to write the RFQ, weight suppliers on evidence rather than price, and check what arrives before you pay for a batch.Read the procurement checklist →

Engineering detail

Detection is not tracking

Detection answers what is in the frame; tracking answers where it went. A system that detects well and tracks badly is unusable, because the operator spends the mission re-gating the target instead of watching it. That is why the tracker is evaluated on occlusion, on scale change across a zoom, and on how it behaves when the target stops moving.

Zoom changes the problem

A tracker that works at 4 mm across a wide field is not the same tracker at 129 mm, where the object fills the frame and small angular errors move it out of view. HITS runs tracking across the zoom range, which is what allows the operator to narrow the field without losing the lock.

Tracking feeds the map

A track without a coordinate is a picture. With the rangefinder in the loop, the payload reports latitude, longitude and altitude for the tracked object, which is what turns an ISR sortie into a product a customer can act on.

Detection and tracking in the video feed

Recorded from the payload video stream rather than a simulator: the boxes are drawn on the payload, so what the operator sees is what the module classified and locked.

Multi-target tracking. Surface contacts detected, classified and tracked in the live feed.

Single-target lock. One target held at long range while the background moves.

Daylight channel, recorded from the payload: the module finds the vehicles in frame, boxes them and holds each lock while the aircraft moves.

The same loop on the thermal channel, where the lock survives on a target that has almost no contrast left in the visible image.

Designed and manufactured by HITS

Optics, servo hardware, stabilisation firmware and the AI tracking module are developed in house, and every unit is assembled, calibrated and environmentally tested on our own line before it ships.