EXPERTISE / ROBOT SOFTWARE

Modern AI Perception

Picking the neural network is the easy part - we make it do real work: We design the whole chain - sensors, network, data, training, tracking, deployment - and run it as an engineering process that gets measurably better every week.

Explore the approach ↓

AI-generated office scene: engineer reviewing perception tracking and plots during a stakeholder video meeting.

01 / THE OPPORTUNITY

Make perception a capability you can improve every week

A promising detector opens the door to useful robot behaviour. Making that capability reliable - and improving it every week - depends on three things:

The network has to fit the hardware. Accuracy, latency and licence trade-offs need to be measured on the target hardware, alongside the rest of the software stack. The best benchmark score is only useful if the robot can run it.
Better data needs a repeatable loop. Consistent labels, separate training and test datasets, and evaluation across lighting, weather and new sites let the team distinguish real progress from a flattering score.
Frame-to-Frame tracking done right. The robot needs stable object identities across cameras, accurate 3D positions and speeds, and rules for when to act. Those capabilities require tracking and task logic beyond the network.

That is an opportunity to build an object detection and tracking pipeline your team can understand, measure and improve as the system encounters new conditions.

02 / HOW WE SOLVE IT

Perception as a 'glass' pipeline

We have built this loop for object detection and tracking on outdoor vehicles, forklifts and stationary cameras. What we bring:

eight-stage perception loop, generated in Gemini in the shared expertise style

Benchmark and select. We train and compare candidate networks on your data and your target hardware, and measure accuracy, latency and licence implications.
Data and labeling pipeline. Recording campaigns with clear scenario coverage; labeling guidelines and class definitions that annotators can actually apply; agreement scoring between labelers; automated dataset assembly with leakage checks, class merging and oversampling of rare classes; model-assisted clean-up of noisy labels.
Training and evaluation you can repeat. Versioned experiments, per-class metrics, confusion matrices, confidence-threshold sweeps - the number you deploy with is the number you measured.
From detections to decisions. Multi-camera fusion, 3D tracking with stable identities across camera hand-overs, speed and trajectory estimation, and the task logic on top (state machines, triggers, alerts) - with tracking metrics (MOTA, identity switches, localisation error) so tuning is quantitative.
Deployment on embedded GPUs. TensorRT / DeepStream pipelines on NVIDIA Jetson, profiled for latency and reliability; the same pipeline replays recorded data for regression testing.
Synthetic data where reality is scarce. Simulation and synthetic data generation to cover rare cases and to test the full vision-to-control loop before hardware exists.

The goal is a perception pipeline with measurable performance on your data and hardware, and a repeatable way to improve it as field conditions change.

03 / IN PRACTICE

In practice: detection and 3D tracking for an outdoor vehicle (USA)

labeled bottle detection in grass, with an enlarged detail of the detected object; user-provided case image

Situation. An outdoor vehicle in the USA needed object detection and live 3D position tracking of objects in the wild.

What we did. We benchmarked several object-detection networks, selected the best trade-off for the target GPU and built the full pipeline: in-field data collection, a labeling operation with guidelines and labeler-agreement tracking, leak-safe dataset assembly, weekly training/evaluation cycles and a multi-camera 3D tracker.

Outcome. Tuning the tracker with proper metrics gave roughly 2× better identity consistency, ~4× fewer identity switches and ~4× lower position error - reproducible across datasets and hardware.

Also: Synthetic data and simulation were then added to test the complete vision-tracking loop.

In practice: camera-based truck tracking for a factory yard (Belgium)

Situation. A factory yard in Belgium used a traffic-sign rule that depended on the site's geometry. The system needed to handle trucks arriving from any road and respond when they started reversing.

What we did. We replaced that rule with a direction-aware tracker. A fixed camera and an NVIDIA Jetson detect and track semi-trucks and switch a traffic sign to red when a truck starts a reversing manoeuvre.

Outcome. Traffic jams at the gate resolved and no more unprotected reversing.

Also: The system runs unattended on a 4G link and records its own validation data.

04 / WHAT YOU GET

What you get

A network benchmark on your data and hardware, with a recommendation you can defend.
A data & labeling pipeline (guidelines, tooling, QA, automated dataset builds) that your team or a labeling vendor can run.
Trained, evaluated models with per-class metrics and a deployment threshold - and the scripts to reproduce them.
Tracking and decision logic with quantitative metrics, plus an embedded deployment on Jetson.
A weekly improvement rhythm: metrics report, decisions on classes/data/model, next iteration.

Engagement: feasibility study (benchmark + data assessment) → development → support for the improvement cycle.