Is the object recognition running on the AI built into the camera, or on the development platform (PC, Embedded Linux devices etc)?

ZED cameras do not embed an AI processor: object recognition, like all the other ZED SDK features, runs on the host device (a PC with an NVIDIA® GPU, an NVIDIA® Jetson™ module, or a ZED Box). The camera streams synchronized left and right images, and the ZED SDK processes them on the host GPU using CUDA and TensorRT.

The ZED SDK Object Detection module combines AI detection with depth information to localize and track objects in 3D:

• Multi-class detection: built-in models detect several object classes (for example persons, vehicles, bags, animals, electronics, fruits and vegetables, sport equipment), with 3D position, 3D bounding box, velocity, and tracking ID.

• Custom detection: you can run your own detector (for example a YOLO-like ONNX model) and let the ZED SDK compute the 3D position and tracking of the detected objects.

• Body Tracking: 3D skeleton detection and tracking of people.

These features are available with all the ZED stereo cameras (ZED 2i, ZED Mini, ZED X, ZED X Mini, ZED X Nano) and with a dual ZED X One Virtual Stereo camera. An NVIDIA® GPU with CUDA support is required; the ZED SDK is free of charge.

More information:

• Object Detection documentation

• Body Tracking documentation

• ZED SDK samples on GitHub