The ZED SDK allows you to add depth, motion sensing and spatial AI to your application. Available as a standalone installer, it includes applications, tools and sample projects with source code.
Blackwell GPUs must use TensorRT 10 and CUDA 12
Pascal GPUs must use TensorRT 8
ZED SDK 5.0 delivers 5× faster depth sensing, 300% lower compute load on Jetson, and improved depth accuracy in challenging environments. It introduces new depth modes powered by TERRA AI, offering options for higher accuracy and faster depth processing to suit different applications. The update also debuts Magellan™, a vision-based localization system providing centimeter-level positioning indoors and outdoors. Additionally, the capture pipeline architecture has been optimized to improve stability, reduce dropped frames, and lower video latency to under 60ms, delivering one of the lowest latencies in the industry.
May 16, 2025
Please note that 5.0.1 and 5.0.0 are not binary compatible. The Python API package will only work with 5.0.1. You should compile it from source if needed to run with 5.0.0.
retrieveObjects and retrieveBodies functions without runtime parameters have been added as well.ENU2Geo and Geo2ENU in sl::Fusion, making it easy to compute latitude-longitude coordinates of a 3D point. See the C++ Geoloc playback sample (convert2Geo function).GEN_3.GEN_3 loop closure detection and map optimization.GEN_3 area mapping file size and saving/loading times.setSVOPosition functions using index or timestamp input. It should now set the expected frame.retrieveImage output when using specific resolutions. The issue could affect grayscale or low-resolution images.isVideoSettingsSupported function with the AEC_AGC_ROI setting that would return invalid results.retrieveObjects and retrieveBodies with runtime parameters is now deprecated. Setting runtime parameters should now be done using the dedicated setters.Mar 19, 2025
NEURAL_LIGHT mode, and updated NEURAL and NEURAL_PLUS. They're faster, more precise, and more robust. The older computer vision modes are now deprecated, the previous NEURAL models can still be accessed for retro compatibility.sl::Camera::read() function that optionally separates the grab() function into two distinct operations: acquiring new images and processing data. This function does not affect existing applications using grab(), it will detect if the read() function was called or not. It enables parallel computing tasks, such as object detection.sl::Resolution(-1,-1) selecting the internal best compromise for both quality and speed.GEN_3, which offers enhanced accuracy and stability.VIDEO_SETTINGS::SCENE_ILLUMINANCE to get the Luminance (LUX) of the scene when using ZED X camerasGEN_2Camera::retrieveImage and Camera::retrieveMeasure for improved performance in GPU asynchronous workflows.blobFromImage, and blobFromImages, for converting images to Deep Learning model tensor inputs.Mat::convertColor, for common color conversions, such as swapping red and blue channels and removing the alpha channel.sl::CameraOneInitParameters::depth_stabilization value set to 30, it provides a more stable depth with minimal motion artifactsCamera::retrieveObjects to Camera::retrieveCustomObjects for custom object detection. The default behavior remains unaffected, but the new method is required when using CustomObjectDetectionRuntimeParameters.CustomObjectDetectionProperties struct:(min|max)_box_(width|height)_meters, to give control to maximum 3D objects dimensionsnative_mapped_class, to allow remapping a custom label to the SDK’s internal SUBCLASS and profit the internal tuningobject_acceleration_preset and max_allowed_acceleration to have better control of the tracked objects' maximum accelerationGEN_2resetPositionalTracking when using Positional Tracking GEN_2read() function for more efficient asynchronous detection.For older releases and changelog, see the ZED SDK release archive.