Embedded vision

Build real-time camera applications at the edge

Combine fine-tuned detection and segmentation, anomaly inference, tracking, regions, and counters in local pipelines. Evaluate each workflow against its target camera, hardware, and latency needs.

Choose the application

Build for the camera task and setting

Industrial inspection

Use anomaly detection for unusual appearance and detection or segmentation for known defects. Connect results to the inspection line's tracking and region logic.

Explore industrial inspection →

People flow and queues

Detect and track people, count crossings, or measure tracked time within a defined region.

Explore queue monitoring →

Produce inspection and counting

Locate and count produce in camera feeds, then evaluate quality workflows against representative crop data.

Explore agriculture →

01 · Edge inference

See an edge prototype running

This demo shows apple detection on a portable Raspberry Pi 5 and Hailo prototype. The recorded output is specific to this device, model, and input.

Use the sample as a starting point for testing a camera, model, and runtime combination on the intended device.

Apple detection on a portable Raspberry Pi 5 and Hailo prototype.

02 · Compare runtimes

Compare tomato and apple detection runs

Sample on-screen FPS readings are about 26 FPS with Hailo HEF and 4 FPS with ONNX for tomato detection, and 14 FPS with Hailo HEF and 3 FPS with ONNX for apple detection. Actual throughput depends on the model, device, input resolution, and runtime.

Tomato detection · Hailo HEF
Tomato detection · ONNX
Apple detection · Hailo HEF
Apple detection · ONNX

These clips show standard color-image detection. For multispectral projects, see Abraia Vision model development and multispectral imaging; validate sensor input and the deployment path for each target setup.

03 · Application integration

Turn a tested workflow into application code

Vision Studio supports data review and workflow prototyping. Use Abraia's Python APIs and command-line tools to build a custom application. Install the package with pip install -U abraia.

Choose runtime support for the target. The package runs compatible ONNX models and supports compiling selected detection or segmentation checkpoints for Hailo. Anomaly review and pipeline inference currently use ONNX.

Validate model, camera resolution, frame rate, latency, and operating conditions on the intended hardware. For visual retrieval, the package also includes image embeddings and semantic search; see retail operations.

View the Python package documentation and examples →

Vision Studio pipeline with strawberry detection, tracking, and a line counter
Studio pipeline · Detection, tracking, and line counting in one workflow.

Train and curate

Prepare datasets, review annotations, and train classification, detection, segmentation, or anomaly models.

Run and compose

Use image processing and image or video inference; compose tracking, region filters, line counts, timing, and ordered pipeline stages.

Deploy to edge

Integrate compatible model formats with the target runtime and validate speed and accuracy on the device.

Plan an edge application

Share the target camera, hardware, model task, and latency needs. We can scope a representative evaluation and development path.