IN Brief:
- Advantech will show a two-node Smart CityMind and EdgeSight Sentinel architecture using NVIDIA Jetson Thor and Jetson AGX Orin.
- Partner demonstrations include existing-CCTV analytics, Ouster 3D LiDAR traffic monitoring, and GPU-based public-space analysis.
- The configurations are exhibition demonstrations at Smart City Expo World Congress on 3–5 November, not confirmed municipal deployments.
Advantech will demonstrate distributed edge artificial intelligence systems for urban applications at Smart City Expo World Congress 2026, combining NVIDIA computing hardware with camera analytics, three-dimensional LiDAR sensing, and autonomous robotics.
The company will exhibit in Hall 3, Booth C161, at Fira Barcelona from 3 to 5 November. Its presentation includes configurations developed with technology partners for public transport, traffic monitoring, commercial spaces, and visual analysis close to the point where data is collected.
To divide continuous visual monitoring from higher-level interpretation, Advantech is pairing Smart CityMind on its NVIDIA Jetson Thor-based AFE-A702 with EdgeSight Sentinel on the Jetson AGX Orin-based AFE-R750. The two devices have distinct processing roles and exchange information within a distributed AI arrangement.
The Thor-based platform acts as a vision-language AI hub intended to interpret information from an urban environment and coordinate work undertaken by connected devices. EdgeSight Sentinel provides continuous visual monitoring at the field node and communicates observations to the higher-level system.
Splitting the functions between nodes can allow local processing to identify relevant events before information is sent elsewhere. Camera systems may produce continuous video streams, but a remote application sometimes needs only selected observations, event notifications, or supporting images rather than every captured frame.
That reduction in transferred data depends on the information required for the task. A traffic count may use compact object records, whereas investigation of an unusual event may need retained video or a more detailed analysis. The software must also establish the timing and origin of observations received from different devices.
The company describes the configuration as using two on-device AI agents, with the central hub directing and the field node carrying out monitoring work. This is an application design, not evidence that the system can autonomously make unrestricted decisions concerning public safety or the control of city infrastructure.
Advantech will also demonstrate a video analytics platform running on Jetson Thor. Image processing can involve identifying objects, following movement, and classifying events, with computing requirements influenced by camera resolution, frame rate, model complexity, and the number of simultaneous feeds.
Advantech and Technis (Videtics) will apply a similar edge-processing approach to CCTV feeds from buses and trains, demonstrating object classification, tracking, and event detection on a TREK-50N equipped with NVIDIA Jetson Orin NX. This configuration is intended to extract useful observations from installed transport cameras rather than requiring an entirely new sensing network.
Camera position, resolution, and optical characteristics still determine what can be inferred from those existing feeds. Poor lighting, obstructed views, or crowded scenes may affect detection and tracking, requiring tests under the conditions expected at each installation.
Tracking also depends on matching an observed object across successive images. A person or vehicle may disappear behind another object and later return to the camera’s field of view; the application must manage that uncertainty rather than assuming every successive detection is correct.
For traffic analysis, Advantech is working with Ouster on a system combining Ouster Gemini and BlueCity 3D LiDAR technologies with the MIC-733 embedded computer. NVIDIA CUDA and TensorRT are used to accelerate selected computational and AI inference functions.
LiDAR measures distances using emitted light and the returned signal, allowing a three-dimensional representation of a monitored area to be assembled. Software can use those measurements to estimate the position and movement of road users, subject to sensor coverage, resolution, environmental conditions, and the algorithms applied.
CUDA provides access to parallel processing on compatible NVIDIA hardware, while TensorRT supports inference optimisation and execution. The contribution of these technologies to overall system performance depends on data preparation, memory transfer, processing resources, and the specific model rather than the accelerator alone.
For shopping centre video analytics, Macnica and Icetana AI will demonstrate NVIDIA RTX PRO 4000 Blackwell technology on Advantech’s MIC-770 and MIC-785 GPU platforms. The higher-performance configuration addresses public-space workloads with processing and deployment requirements that differ from the Jetson Orin NX transport installation.
Advantech will also show a self-driving robot built around the MIC-732 and a 64GB NVIDIA Jetson AGX Orin module. Its navigation software must interpret sensor observations, identify obstacles, estimate position, and issue movement commands within the machine’s physical limits. Before deployment in environments shared with people, engineers would need to validate obstacle detection, localisation, stopping behaviour, and responses to sensor faults under varied lighting, obstructions, and unexpected movement. The exhibition demonstration does not establish accuracy or dependable operation across those conditions.
Advantech’s October development is the integration of its embedded computing platforms with partners’ sensing and analytics technologies into specific configurations for the Barcelona demonstrations, rather than the launch of a new processor or computer family. The company has not identified completed city deployments, disclosed new municipal contracts, or published independently verified detection accuracy and latency results for the full configurations.
The demonstrations show how computing resources can be distributed between sensing devices, field controllers, and higher-level processing nodes. Any production installation would still need evaluation of data interfaces, power and cooling, communications reliability, model performance, and application-specific safety requirements before the proposed architecture could be validated.
