> For clean Markdown of any page, append .md to the page URL. > For a complete documentation index, see https://docs.stereolabs.com/llms.txt. > For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.stereolabs.com/_mcp/server. # Positional Tracking Modes The StereoLabs Positional Tracking module uses visual-inertial SLAM (VSLAM) to deliver precise, real-time estimates of the camera's 3D position and orientation. By fusing stereo vision with IMU data, it builds and refines a 3D map of the environment while simultaneously tracking motion within it. This approach ensures accurate, drift-minimized localization, even in dynamic or visually challenging conditions, making it suitable for demanding robotics, AR/VR, and autonomous navigation applications. The ZED SDK includes multiple generations of positional tracking algorithms (`GEN_1` and `GEN_3` 1), each designed to meet different performance and accuracy requirements. The positional tracking module supports two operating localization strategies: * **VIO (Visual-Inertial Odometry)**, which performs pure localization without using a prior map. * **Relocalization**, which performs localization within an area map, allowing the system to recognize known environments and maintain consistent tracking across sessions. These modes and localization strategies allow you to choose the best balance between computational load, precision, and robustness for your specific use case. 1 *The `GEN_2` positional tracking generation has been deprecated and will be removed in a future release*. ## `GEN_1` mode `GEN_1` implements a dense VSLAM approach that leverages depth data from the stereo camera to estimate motion. It generates a dense representation directly from depth information rather than sparse keypoints, enabling more stable and drift-resistant tracking in low-texture or feature-poor areas. The fusion of stereo vision and IMU data improves robustness against motion blur and rapid movements. Originally designed for AR/VR applications, `GEN_1` delivers smooth and consistent positional tracking, especially during fast or complex movements. > **Note** > > `GEN_1` is optimized primarily for VIO (Visual-Inertial Odometry) mode. While it excels at real-time localization and smooth tracking, its dense VSLAM architecture was not designed with loop closure or area map relocalization in mind. For applications requiring persistent mapping or multi-session relocalization, `GEN_3` is the recommended choice. ### `GEN_1` Load Performances | Depth Mode | Max FPS | CPU (%) | GPU (%) | | ------------ | ---------------------------------------------------------------------------------------------- | ------- | ------- | | Neural Light | [Depth Compute FPS](/docs/development/zed-sdk/modules/depth-sensing/depth-modes/#neural-light) | 13 | 77 | | Neural | [Depth Compute FPS](/docs/development/zed-sdk/modules/depth-sensing/depth-modes/#neural) | 8 | 80 | | Neural Plus | [Depth Compute FPS](/docs/development/zed-sdk/modules/depth-sensing/depth-modes/#neural-plus) | 4 | 94 | > **Note** > > Performance obtained with **ZED SDK v5.1.1**, **ZED X Driver v1.3.2**, and ZED X camera using the positional tracking [sample](https://github.com/stereolabs/zed-sdk/tree/master/positional%20tracking/positional%20tracking/cpp) available on GitHub. Because `GEN_1` is depth-dependent, performance varies based on the chosen depth mode. More information on depth modes is available on this [page](/docs/development/zed-sdk/modules/depth-sensing/depth-modes). ### `GEN_1` Accuracy Performances | Depth Mode | Mean APE \[m] | Max APE \[m] | Specific Scenario Dataset | | ----------------------------------- | ----------------- | ----------------- | ---------------------------------------------------- | | Neural Light / Neural / Neural Plus | 0.8 / 0.8 / 0.62 | 1.8 / 1.78 / 1.47 | Indoor warehouse environments with reflective lights | | Neural Light / Neural / Neural Plus | 3.38 / 3.9 / 2.45 | 7.76 / 8.27 / 5.7 | Structured outdoor sceneries during the day | > **Note** > > Performance obtained with **ZED SDK v5.1.1**, **ZED X Driver v1.3.2**, and ZED X camera using the positional tracking [sample](https://github.com/stereolabs/zed-sdk/tree/master/positional%20tracking/positional%20tracking/cpp) available on GitHub. Each value represents the average result from a series of real-world tests collected under a specific scenario dataset (e.g., indoor, outdoor). The average test sequence spans approximately 400 m. ## `GEN_3` mode `GEN_3` introduces a scalable, feature-based VSLAM pipeline built for robustness and precision. By extracting and tracking high-quality visual features (rather than relying on `GEN_1`’s dense mapping), it maintains an accurate, lightweight map with strong loop closure and global optimization capabilities. This minimizes drift, improves long-term consistency, and delivers fast and reliable relocalization. Combined with visual-inertial fusion, `GEN_3` adapts seamlessly to large, dynamic, or revisited environments, ideal for advanced robotics and autonomous navigation applications. > **Note** > > `GEN_3` is recommended for both the VIO and relocalization localization strategies. ### `GEN_3` Load Performances | Platform | Max FPS | CPU (%) | GPU (%) | | ------------------ | ------------ | ------- | ------- | | Jetson™ Orin NX 16 | up to 80 FPS | 13 | 44 | > **Note** > > Performance obtained with **ZED SDK v5.1.1**, **ZED X Driver v1.3.2**, and ZED X camera using the positional tracking [sample](https://github.com/stereolabs/zed-sdk/tree/master/positional%20tracking/positional%20tracking/cpp) available on GitHub. ### `GEN_3` Accuracy Performances | Mode | Mean APE \[m] | Max APE \[m] | Specific Scenario Dataset | | ---- | ------------- | ------------ | ---------------------------------------------------- | | VIO | 0.56 | 1.2 | Indoor warehouse environments with reflective lights | | VIO | 0.29 | 0.58 | Structured outdoor sceneries during the day | > **Note** > > Performance obtained with **ZED SDK v5.1.1**, **ZED X Driver v1.3.2**, and ZED X camera using the positional tracking [sample](https://github.com/stereolabs/zed-sdk/tree/master/positional%20tracking/positional%20tracking/cpp) available on GitHub. Each value represents the average result from a series of real-world tests collected under a specific scenario dataset (e.g., indoor, outdoor). The average test sequence spans approximately 400 m. ## Positional Tracking Modes Comparison | Positional Tracking Generation | Benefits | Limitations | | ------------------------------ | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | | GEN\_1 | * Well-suited for environments with limited visual information (low texture, low light, repetitive structures). * Reliable for open-field outdoor robotics or inspection tasks. * Optimized for AR/VR applications. | - Depends heavily on depth information, resulting in higher computational requirements. - Provides fewer opportunities for loop closure, making it less effective in feature-rich indoor environments. - Only suited for VIO localization strategy, not recommended for relocalization. | | GEN\_3 | * Designed for feature-rich environments such as indoor facilities, warehouses, offices, and structured outdoor scenes. * Offers advanced loop closure and relocalization, making it ideal for area mapping and long-term navigation. * Provides a lightweight and efficient tracking pipeline that scales well to large and complex spaces. * Well suited for both VIO and Relocalization localization strategies, without the need for depth. | - May be less robust in severely degraded visual conditions where very few features are available. |