Vision AI

Our experience comes from industrial machine vision and vehicle perception. The same methods a vehicle uses to perceive its environment also analyse aerial images and recognise the poses and movements of people.

Measuring where measurement is possible

For us, classic image processing and machine learning are not opposites. Where a feature can be measured, such as a length, a position, or a shape, we measure. Where appearances are diverse and hard to capture in rules, we use neural networks. Many of our systems combine both: the network finds and outlines the object, the measurement evaluates it.

Four perspectives, the same methods

Stable over time

A detection in a single image is rarely enough. We consistently track detected objects across the image sequence. This makes the detection stable, reduces the number of false alarms, and additionally provides direction of movement and speed. In person detection, the goal is high detection performance combined with a low false alarm rate.

Where the training data comes from

Machine learning needs examples, specifically from the later area of use. We collect, organise, and version image data and annotate it with our own tools; for aerial images, this is Acotate 2.0 AERO. Automatic pre-annotation speeds up the work; people correct and approve. Where real recordings are missing or too rare, we add synthetically generated images.

Image and language

Vision-language models evaluate an image together with text. This makes it possible to solve tasks where pure image recognition fails: identifying a weathered traffic sign with the help of the standard description, or classifying a document page by its function. How we use this is described under Multimodal AI.