Executive Summary
Amid high Security Operations Center (SOC) operator fatigue and burnout, mobile cameras are becoming increasingly commonplace. Today’s CCTV systems assume cameras are static and that their locations are known. But as robots, drones, body-worn and vehicle-mounted cameras become ubiquitous, this assumption is no longer valid. SOC operators will need to understand not only what each camera sees but also where it is in the world. Recent advances in 3D reconstruction models make it possible to reconstruct environments and estimate camera locations from ordinary video automatically. We show how state-of-the-art AI models can be integrated into existing modern video management systems, such as Milestone XProtect, providing SOC operators with visualizations of camera pose and orientation.
A Milestone Smart Client view showing two video streams at the top and a visualization showing each camera’s pose at the bottom.
As organizations add more cameras, Security Operations Center (SOC) operators have to stitch more viewpoints into one picture of the site. Their cognitive load rises with every extra feed. The rise of mobile cameras, which are becoming increasingly commonplace [1], further compounds this challenge by breaking the long-held assumption that cameras are fixed in place. Security robots are starting to patrol autonomously, drones are being used for perimeter monitoring, and body-worn and vehicle-mounted cameras are already universal. Together, these trends increase the complexity of monitoring tasks at a time when analyst fatigue and burnout have already been identified as key concerns for the security industry [2].
This means that SOC operators can no longer rely on a fixed map. Fortunately, recent advances in 3D reconstruction models make it possible to infer camera positions and orientations from ordinary video streams. This enables video management systems to become spatially aware, without requiring dedicated localization hardware or manual calibration. This can reduce the mental load required from SOC operators to determine the location and viewpoint of a particular camera at a particular time.
To illustrate how this capability can help SOC operators, we use the dataset published by Sanchez-Iborra et al. [3], featuring a stationary CCTV camera and a vehicle-mounted dashboard camera. By applying a 3D reconstruction model to this video data, which can infer relative camera poses as a byproduct of the 3D reconstruction process, we can automatically recover the relative camera positions at each point in time. To present this information in a format that is as easy and quick as possible for SOC operators to understand, we integrate this pose data directly into the VMS.
The two videos used in our example are: the left is the dashboard camera, and the right is the CCTV camera, both taken from Sanchez-Iborra et al. [3].
Reconstructing a 3D scene from images alone has long been a goal for researchers, and models have been able to recreate 3D scenes accurately for several years. However, until recently, the performance, latency, and compute requirements of these models hindered their use in production settings. Recent feed-forward models can reconstruct scene geometry in seconds or less [3], while newer models such as VGGT-Ω [4] further improve reconstruction accuracy and computational efficiency, making practical deployment increasingly feasible.
3D reconstruction models take images or videos as input and predict key 3D attributes of a scene, including camera parameters, point maps, and depth maps. Using modern transformer architectures and training methods, reconstruction can occur in a single forward pass, meaning the model can directly predict point clouds or even Gaussian splats. VGGT [5] was one of the first models to directly infer point maps, depth maps, camera parameters, and matching points, all in a single forward pass.
However, this model was very compute-hungry. For 200 frames, it required 40.63 GB of memory on an H100 GPU. Its successor model, VGGT-Omega, improves performance while also reducing compute: it uses only 20.82 GB of memory for 200 frames on an A100 GPU.
We see the same pattern with other approaches: Depth-Anything-3 [6] outperforms Depth-Anything-2 [7] while also reducing compute. This enables real-world use of these models.
Using VGGT-Omega, we reconstruct the shared 3D scene from the two videos, thereby also estimating the relative positions of both cameras. Although the two cameras observe each other only during part of the sequence, and much of each video captures distinct, non-overlapping perspectives of the scene, the model successfully recovers the scene and provides a stable estimate of the relative camera position and orientation.
To ensure this camera pose information is useful, it must be integrated directly into SOC operators' existing workflows. To demonstrate this, we integrated our camera localization pipeline with Milestone XProtect. XProtect is the on-premises platform that manages camera onboarding, live and recorded video streams, and access control across an organization’s security infrastructure. It exposes cameras and their streams to third-party analytics applications, allowing analytics to be sent back to operators in the system they use day to day. Integrating with XProtect allows the camera pose information to be presented natively inside the VMS, next to normal video feeds.
Furthermore, using an established VMS is critical when creating AI applications for security systems, as organizations can build on a secure, tested foundation and avoid reinventing the wheel when handling high-volume data such as video streams. An existing VMS like Milestone XProtect provides a robust, secure, and scalable framework designed to manage thousands of camera streams simultaneously, ensuring reliability that is difficult to replicate in custom-built systems. Additionally, leading platforms are architected to be extensible, exposing APIs and integration platforms that allow the addition of AI analytics. By building within this ecosystem, we can focus our engineering efforts where it really counts, such as estimation of camera pose, while offloading the undifferentiated work of stream ingestion, authentication, user management, and video archiving to a platform already optimized for these tasks.
In particular, Milestone provides the Milestone AI Bridge, which follows their philosophy of providing flexibility to external integrations. The AI Bridge is a containerized middleware layer that lets external AI applications interoperate with XProtect without using Milestone’s protocols directly. This is needed because XProtect VMS runs on Windows machines, while most AI applications run on Linux machines. This allows software from these two different ecosystems to communicate. This also lets us dynamically scale the video-processing backend based on needs. This is crucial because it enables cost-effective use of GPU-accelerated resources.
To capitalize on these advantages, we designed a backend to estimate camera pose that prioritizes efficiency and scalability. The backend is hosted on Google Cloud Platform, which ensures enterprise-grade security and robust compliance standards. A Cloud Run Service maintains a persistent connection to the two live camera streams and accumulates the incoming frames in the Google Cloud Storage (GCS) bucket. Once enough frames have accumulated, it starts a Cloud Run Job which deploys a container attached to a GPU to perform the forward pass with VGGT-Omega. This setup ensures GPU time is used only when necessary, allowing the most expensive part of the backend to scale to zero. Once the flythrough video has been generated and written to the GCS bucket, the Cloud Run Job finishes.
Architecture diagram of an example backend on Google Cloud.
The AI Bridge handles all communication between the backend and Milestone XProtect, brokering and authenticating all traffic within the Milestone ecosystem. For the two inbound camera streams, the service resolves each camera’s RTSP URL via a GraphQL query against the Bridge’s web service, then reads frames using OpenCV. For the outbound streams, the service registers itself as an application and video topic with the AI Bridge using GraphQL, resolves the gRPC endpoint for that topic, and opens a client-streaming gRPC call over which it sends JPEG-encoded frames of the flythrough video. The AI Bridge then republishes that stream to XProtect as a viewable virtual camera, so operators see the inferred camera positions alongside their normal live feeds.
Video management systems have traditionally assumed that cameras occupy known, fixed locations. As surveillance evolves to include robots, drones, vehicle and body-worn cameras, that assumption no longer holds. In the future, SOC operators will need to understand both what a camera sees and where it is in the environment, and VMS systems will need to supply that information.
Recent advances in 3D reconstruction models make this more practical than ever before. Modern models such as VGGT-Omega can reconstruct environments and estimate camera trajectories directly from ordinary video with substantially lower latency and computational cost than previous approaches.
Our example application illustrates how these capabilities can be integrated into Milestone XProtect using the AI Bridge and cloud-based inference, allowing reconstructed scenes to appear directly inside the operator's existing workflow. While this is only one example application, the same underlying capability opens the door to implementing state-of-the-art AI models in existing video management systems.
[1] Research and Markets ltd, “Research And Markets,” researchmarkets, 2026. https://www.researchandmarkets.com/reports/5303706/mobile-video-surveillance-global-strategic
[2] S. Tariq, Mohan Baruwal Chhetri, Surya Nepal, and C. Paris, “Alert Fatigue in Security Operations Centres: Research Challenges and Opportunities,” ACM Computing Surveys, vol. 57, no. 9, Mar. 2025, doi: 10.1145/3723158.
[3] R. Sanchez-Iborra et al., “Dataset for multi-perspective traffic video analysis,” Scientific Data, vol. 13, no. 1, Feb. 2026, doi: 10.1038/s41597-026-06907-y.
[4] J. Wang, M. Chen, S. Zhang, N. Karaev, J. Schönberger, P. Labatut, et al., “VGGT-Ω,” arXiv preprint arXiv:2605.15195, 2026.
[5] J. Wang, M. Chen, N. Karaev, A. Vedaldi, C. Rupprecht, and D. Novotny, “VGGT: Visual Geometry Grounded Transformer,” arXiv.org, 2025. https://arxiv.org/abs/2503.11651 (accessed Aug. 19, 2026).
[6]H. Lin et al., “Depth Anything 3: Recovering the Visual Space from Any Views,” arXiv.org, 2025. https://arxiv.org/abs/2511.10647 (accessed Aug. 19, 2026).
[7] L. Yang et al., “Depth Anything V2,” 2024. Accessed: Aug. 19, 2026. [Online]. Available: https://arxiv.org/abs/2406.09414