Cities worldwide are investing in smarter infrastructure and data-driven solutions to improve safety, transportation, and quality of life for their residents. For example, video cameras installed in public spaces detect and count people in crowded areas to identify situational risks in real time – enabling faster response to emergencies. On city streets, the same technology helps monitor traffic and public transportation for safe and efficient travel. In both use cases, camera and computer vision technology can provide trend data for infrastructure and transportation planning and services.
Real-time people counting and situational awareness require that AI machine vision be processed at the edge, versus sending video data to the cloud. To do so demands low-latency and high-performance compute, packaged in a small and rugged form factor.
Open computer-on-module (COM) standards such as SMARC are ideal for integrating edge AI processing into compact camera systems, as they offer high computing power combined with a small footprint and high energy efficiency. And commercial off-the-shelf (COTS) COM-based designs support long device lifecycles, with common carrier boards for flexible designs, which enable faster time-to-market.
Powerful SMARC Module for AI Processing at the Edge
Developers designing smart city surveillance applications, like people counting, can rely on the SECO SOM-SMARC-MX95 to streamline design and deployment. The compact SMARC (Smart Mobility ARChitecture) 82 mm × 50 mm form factor enables camera system integration even in the most space-constrained situations. The SMARC module includes the following key features:
- CPU: NXP i.MX 95 with 6x Arm Cortex-A55 cores, 1x Cortex-M7 core, and 1x Cortex-M33 core.
- NPU: NXP eIQ Neutron for execution of AI models directly at the edge.
- Modern graphics support: 2D/3D GPU.
- Fast memory: Up to 16GB LPDDR5 6.4GT/s (32-bit).
- Extensive connectivity: 2x GbE, support for 1x 10 GbE via XGMII, optional Wi-Fi + BT/BLE module.
- Operating temperatures: 0 to +60 °C for commercial and -40 to +85 °C for industrial applications.
The numerous SOM-SMARC-MX95 interfaces – MIPI CSI-2 camera interfaces, HDMI/LVDS display outputs, PCIe Gen3 x1, USB 2.0 and USB 3.0, and Gigabit Ethernet in particular – facilitate the connection of multiple sensors and networking into larger, distributed surveillance systems.
Setting up People Counting Systems
With the SOM-SMARC-MX95, developers can easily set up people counting systems: one or more cameras and other peripheral devices are connected to a carrier board on which the module is integrated and works as the central compute core, performing AI-based evaluation directly at the edge.
Several camera nodes can be combined to form a larger networked system. Depending on the security and data protection requirements of the application, video data or compressed inference and counting data can also be transmitted from the edge for processing in the cloud.
Together, the SOM-SMARC-MX95 and Clea OS, SECO’s cross-platform embedded Linux based on Yocto Project, form a pre-integrated hardware/software stack that greatly simplifies the development of edge AI-based camera systems. The SOM-SMARC-MX95 with the NXP i.MX 95 processor provides AI-accelerated computing power and extensive camera connectivity for edge vision applications. Clea OS delivers a preconfigured, Yocto-based Linux operating system with configurations specific to the SOM-SMARC-MX95 – eliminating the effort typically required for low-level system initialization and configuration.
In this combination, Clea OS enables end-to-end camera pipelines – from image capture to AI-accelerated inference to people counting. And the optimized AI frameworks support fast and reliable implementation of edge AI applications.
YOLOX: A Flexible Platform for Computer Vision
Besides powerful hardware, a developer-friendly operating system, and the right AI processing software are also crucial for successful edge AI applications. YOLO, for example, is a model architecture and a family of deep learning algorithms, ideal for object recognition. It is particularly suitable for real-time object recognition and is used in numerous smart city and industrial applications. Based on this architecture, there are different models available, like YOLOv3, YOLOv5, or YOLOX, a modern, powerful variant. After optimization, YOLOX is ideal for use on resource-constrained edge devices such as the SOM-SMARC-MX95, where it enables efficient inference with low energy and memory requirements.
SECO has successfully tested this SMARC module with YOLOX, using an object recognition and people counting model based on a LiteRT framework. The workload described in the SECO App Hub includes camera data capture, preprocessing, AI inference on the eIQ Neutron NPU, and downstream evaluation and output of the counting results – demonstrating a practical edge AI pipeline for people counting applications.
For this workload, CPU-based inference shows clear performance characteristics. Processing achieves a low latency of 411.48 ms using 21.9 MB of memory, which proves the basic functionality of the application. At the same time, it highlights the optimization potential offered by dedicated AI accelerators like the NXP eIQ Neutron NPU, making the SOM-SMARC-MX95 an even more promising platform for people counting workloads.
Getting Started with SECO’s Development Platform
Designed to streamline and expedite edge computing implementations, the DEV-KIT-SMARC makes it easy for developers to get started on SMARC-based projects. The kit includes a development platform that simplifies working with SMARC modules and is ideal for quickly deploying a people tracking application.
The development kit includes a comprehensive set of ports for networking, video, and audio, along with cables, storage, and many other interfaces. With flexibility in development across SECO’s range of SMARC offerings in mind, the SMARC card must be ordered separately.
For people counting applications, with two CSI input interfaces, multiple cameras in proximity can be directly connected to the carrier board, and two dual RJ-45 Gigabit Ethernet ports enable the networking of individual camera nodes or connecting to surveillance systems. USB sockets and an HDMI port support the integration of displays and keyboards for local HMI. This is complemented by flexible power supply options that allow stable operation, both in the lab and in realistic demonstration setups.
Conclusion
Counting people and monitoring crowded areas in real time is essential for city security and safety around the globe. But successful deployments demand high-performance compute to support computer vision and AI inferencing at the edge. The SOM-SMARC-MX95 provides the computing power, compact dimensions, extensive camera support, multiple I/O, and network interfaces required to develop people counting solutions.
On top of the hardware, SECO shows how YOLOX AI models can be deployed directly at the edge—powered by the SECO Clea OS, ready-to-use Yocto operating system. Finally, embedded system designers have a streamlined path to evaluating and deploying people counting platforms with SECO’s versatile development kit.
Get in contact with SECO experts now and benefit from innovative COM solutions.