What Is Computer Vision?

Computer vision is the field of artificial intelligence that turns pixels into information a machine can act on. A camera supplies the image; a computer vision model works out what is in it, where each object sits, what shape it has and how it is moving. The output is structured data, such as a label, a bounding box, a mask, a count or a measurement, that software can log, display or use to drive a decision.

The field is older than deep learning. Early systems relied on hand-designed features such as edges, corners and gradient patterns, and many measurement and gauging tasks still work that way. The shift came in 2012, when AlexNet, a convolutional neural network trained on GPUs, won the ImageNet image classification challenge by a wide margin. Since then most computer vision has been built on learned models: convolutional networks first, and more recently vision transformers and vision language models.

So, what is computer vision in practice? It is the camera, the model and the computer that runs it, working together to answer a visual question fast enough to be useful.

Whatever the model, a computer vision system follows the same basic pipeline from sensor to decision. The stages matter because a weak link anywhere, from poor lighting to a slow interface, limits the whole system.

Most applications combine a handful of standard tasks. Each produces a different kind of output and uses a different family of models.

Real systems chain these together. A traffic camera detects vehicles, tracks each object across frames and reads the plate; a safety system detects people and runs pose estimation to spot a fall.

The two terms overlap and are often used interchangeably. Computer vision is the broad scientific field. Machine vision is the industrial application of it: cameras, lighting and software on a production line that inspect, measure and guide, usually with fixed positions, controlled lighting and strict timing tied to a PLC.

In practice the difference shows in the setting and the output. Machine vision works with fixed cameras, controlled lighting and known parts, and its answer is a pass or fail, a measurement or a robot coordinate, delivered in milliseconds. Computer vision handles open scenes, such as roads, shops and warehouses, and returns labels, boxes, masks and tracks that other software uses. Deep learning has narrowed the gap, because the same detection and segmentation models now run in both, often on the same edge computer.

Robots, autonomous mobile robots and vehicles use computer vision to understand space as well as objects. Depth from stereo cameras lets a machine judge distance to an obstacle or a shelf. Visual SLAM (simultaneous localisation and mapping) builds a map from camera motion and places the robot in it. Grasp detection finds where a gripper can hold an item, and segmentation separates the floor from people and pallets.

These machines usually carry several cameras at once, often on long, vibration-resistant links such as GMSL cameras, with hardware sync so frames from each sensor line up. Vision is then fused with lidar, IMU and wheel odometry for a more reliable picture than any one sensor gives.

Training a vision model normally happens on GPU servers or in the cloud. Running it is another matter. Video is heavy, decisions are often needed in tens of milliseconds, and many sites cannot send footage off the premises, so most industrial computer vision runs at the edge, on a computer close to the cameras.

We build edge AI computers for computer vision projects with the GPU or accelerator, camera interfaces and operating system image fitted and tested as standard. Send us your camera count, resolution and model, and we will size the system to match.