Tracking
Knowing an object is there once isn’t enough. Autonomous systems need the same vehicle, pedestrian or cyclist to stay the same object frame after frame — with a prediction of where it moves next. Tracking is Provizio’s object-tracking module: it links detections into persistent objects, gives each a stable ID, and carries that identity across time and across radars.
Stable IDs
frame-to-frame continuity
Velocity-aware
Doppler-informed prediction
All-weather
fog · dust · darkness
On-edge
no external compute
Applications
Where the planner needs object continuity, not just per-frame detections.
Automotive
Vehicle, pedestrian and cyclist tracking for ADAS and autonomous driving — AEB target selection, forward-collision warning, adaptive-cruise gap control and lane-change risk. Each one needs a stable track ID from frame to frame, not a fresh detection every time.
Industrial
Worker safety around heavy machinery, and autonomous yard, mine and agriculture vehicles. Tracking gives safety systems the persistence to reason about an approaching person across many frames, instead of reacting to a single detection.
Smart city
Smart-city and infrastructure perception — monitoring traffic on highways, junctions and urban streets. Tracking gives every vehicle a persistent ID for counting, speed and trajectory analysis, in any weather and without cameras.
Surround
On a platform with front, side, rear and corner radars, the multi-radar fusion tracker preserves each object’s identity as it crosses from one sensor’s field of view into the next — handing the track over seamlessly, so nothing is dropped or double-counted.
How it works
From a fresh cloud of detections to objects with a stable identity.
STEP 01
Match each detection to its object
Every frame, incoming detections are matched to the objects already being tracked. Tracking weighs how close each detection is and whether its measured velocity fits, to decide which detections belong to which existing object.
STEP 02
Predict where each object goes next
Radar measures each object’s speed directly from the Doppler shift. Tracking uses that velocity — not just past positions — to predict where every object will be in the next frame, so it stays locked on even through quick changes of speed.
STEP 03
Hold a stable ID and reject clutter
Each object keeps one persistent ID for as long as it is in view. New objects have to be confirmed over several frames before they count, which filters out fleeting clutter and multipath reflections before they reach the planner.
STEP 04
Carry identity across radars
For multi-radar setups, a fusion layer lifts every track into one shared map and hands an object from one radar’s field of view to the next — same ID, same velocity, same history — so a vehicle crossing the gap between sensors is never counted twice.
Featured deployment
Tracking every vehicle on a live dual carriageway.
Mounted on an overpass, a Provizio-enabled radar tracks every vehicle on a busy dual carriageway — giving each a persistent ID and class, and holding it from the moment it enters the field of view until it leaves. That turns raw traffic into per-vehicle counts, speeds and trajectories for smart-city and highway operators — in any weather, day or night, without cameras. It’s the same multi-object tracker that runs in-vehicle, here pointed at the road.
- Overpass / smart-city deployment
- Persistent track ID + class per vehicle
- Per-vehicle count, speed and trajectory
- All-weather, day or night, camera-free
Variants
From one radar to a full surround suite.
Tracking scales from a single sensor to a multi-radar platform — the single-radar tracker, and a fusion layer that extends it across sensors.
SINGLE-RADAR
Per-sensor object tracking
Probabilistic data association and a velocity-aware motion model maintain a stable ID for every object, running in real time on the radar’s onboard compute.
When to use: Any single-radar perception build
MULTI-RADAR
Fusion hand-off tracker
Extends tracking across a radar suite — lifts tracks into one shared map and carries a single global ID as objects pass between sensor fields. The continuity layer behind URMAP 360 surround perception.
When to use: Surround / 360° multi-radar platforms
Integration
Runs on the radar, ships over DDS
Tracking is a microservice in the 5D Perception stack, running on the radar’s onboard edge compute. It consumes detections and class labels from the perception stack and ego-motion from Radar Odometry, and publishes tracked objects — each with a stable ID, position, velocity and heading — over DDS at the radar’s frame rate.
- Runs on the radar’s onboard edge compute
- DDS output: track ID + kinematics
- Consumes URMAP detections + Classification labels
- Doppler-aware motion prediction
- Confirmation logic filters clutter and multipath
- Multi-radar fusion hand-off
Questions, answered
Tracking FAQ
Per-frame radar detections are not enough for autonomy — a pedestrian crossing the road needs to stay the SAME tracked pedestrian frame after frame, not a fresh detection each time. Tracking links detections into persistent objects, gives each a stable ID, predicts where it moves next, and filters out clutter. That track ID is what downstream planners reason about.
It matches each frame’s detections to the objects already being tracked using probabilistic data association — weighing position and Doppler-measured velocity — and a velocity-aware motion model to predict each object’s next position. New objects must be confirmed over several frames before they count, which keeps IDs stable and rejects fleeting clutter.
A multi-radar fusion layer lifts every track into one shared map and hands objects across the gap between sensor fields, giving each a single global ID so a vehicle is never double-counted. It is decentralised — radars cooperate peer-to-peer over DDS — and is the continuity layer behind URMAP 360 surround perception.
On the radar’s onboard edge compute. Tracked objects (ID, class, position, velocity, heading) publish over DDS at the radar’s frame rate, alongside the detection, classification and odometry streams.
From detections to objects you can plan around.
If you need stable, predictable track IDs feeding your planner — on a single radar or across a full surround suite — Tracking is the layer for that.