The SECO SOM-COMe-BT6-RK3588 module presents a powerful embedded computing platform for real-time machine vision at the edge to greatly improve the reliability of security camera systems through built-in AI inference that boosts threat detection and eases the burden on human operators.
Camera monitoring stations form an essential part of modern life, serving everything from office buildings to hospitals, banks, and museums. Within these stations, security operators scan multiple video feeds for anomalies and rush to investigate threats as they arise. But whether screens are filled with crowds or remain empty for an entire shift, sustained hypervigilance can lead to operator fatigue, where mental and physical exhaustion compromise effectiveness.
For example, repeated false positives can cause desensitization to alerts and lead to real threats being ignored. Similarly, monotony induces lapses in concentration where initial suspicious activities are missed entirely. And by the time these develop into real threats, it may be too late for limited security teams to catch up, especially in large buildings or busy settings.
To ease cognitive burden, artificial intelligence vision (AI vision) can automate highlighting both active and potential threats in real time across multiple video feeds. Acting as a force multiplier, AI vision helps human operators focus on relevant areas when things are quiet or prioritize effort during busier shifts. Combining AI with camera systems solves a pressing human-centric security problem, but direct integration requires powerful edge-based computing for maximum system autonomy and real-time inference.
How AI Vision at the Edge Improves Surveillance
AI video analysis offers two main approaches: edge-based processing and processing in the cloud. While the latter can use abundant computing resources to perform the most advanced analysis available, it comes with drawbacks that can compromise security operations.
The main drawback is latency – the time between the video capture and the resulting decisions. Traditionally, latency refers to the time it takes for video data to be sent to the cloud, analyzed, and then returned for AI inference or a combined feed. Given the data volumes for high-definition video, this makes cloud-based video processing unwise for achieving predictable real-time performance, which is essential in security applications as it maximizes the response window for operators. The problem significantly expands with the number of cameras and applications requiring cross-correlation of events across multiple feeds.
Edge-based processing solves this issue as AI inference is done close to the source for minimal latency. Local, edge-based processing is not impacted by external network reliability or job queuing in third-party data centers. Closed, localized security networks improve overall system reliability and reduce the risk of hacking. Keeping all video data within a local network can eliminate ongoing data transfer and cloud-based service costs, making edge-based inference attractive to AI system integrators.
Despite the resource limitations of edge hardware, many sophisticated AI vision models are readily available for deployment. Example workloads for security applications include:
- Facial recognition facilitates approved personnel entering restricted areas or flags unrecognized people.
- Re-identification (Re-ID) tracks individuals across different images and separate camera feeds so suspicious parties are not lost in crowds or over long distances.
- Object detection informs operators of potential threats by flagging items like weapons, stolen items, or high-visibility vests (a disguise that enables actions to go unquestioned.)
- Body movement tracking examines a person’s gestures to flag aggressive or suspicious behavior, such as one focusing on security cameras.
A CoM-Based Approach to Integrated AI
When considering edge platforms for AI vision, the SOM-COMe-BT6-RK3588 computer-on-module (CoM) is an innovative solution. Unlike most COM Express modules, which typically feature x86 architectures, the SOM-COMe-BT6-RK3588 combines an Arm-based Rockchip RK3588 multi-core processing system-on-a-chip (SoC) with a separate Axelera AI Metis artificial intelligence processing unit (AIPU) to enable advanced AI vision within the narrow power envelope defined by the COM Express open standard.
Much of the module’s AI processing capability is owed to the Metis AIPU, a dedicated integrated circuit which offers power efficiency of around 15 TOPS per watt for up to 120 TOPS of AI performance as part of the complete SOM-COMe-BT6-RK3588 platform. This supports edge-based video inference across multiple camera feeds for comprehensive real-time threat detection.
These processing architectures are supported by up to 32 GB of memory for the CPU and a dedicated 2 GB for the AIPU, providing ample resources for video data management during AI analysis. Key camera interfaces include two MIPI-CSI 2-lane connectors and several USB ports, with a single Ethernet port for interfacing with local security networks.
By integrating edge computing via an open-standard commercial-off-the-shelf (COTS) CoM, developers also gain the benefits of accelerated development, long lifetime support without vendor lock-in, and flexible, compact integration via a carrier board. Once designed, this carrier can then serve multiple product generations for faster time-to-market as security applications evolve.
Implementing Real-Time AI Vision with Metis
The Axelera AI Voyager SDK offers a user-friendly, end-to-end integrated software stack that simplifies AI vision model deployment on the Metis AIPU. Whether using proprietary models or popular development frameworks like YOLO and ResNet, Voyager SDK automatically compiles and optimizes the AI pipelines for easier setup of real-time inference. For more information, check out the Axelera AI Voyager SDK GitHub repository.
To complete the SOM-COMe-BT6-RK3588’s software ecosystem, SECO’s Clea OS presents a flexible, Yocto-based platform that engineers can customize to meet the specific demands of AI-enabled security applications. Its modular nature accelerates low-level development and streamlines access to the SOM-COMe-BT6-RK3588’s many on-board processing cores, whether the CoM is integrated into an individual camera for endpoint inferencing or supporting several nearby cameras before relaying analyzed feeds to the main security station.
A Streamlined Future for AI-Driven Security Cameras
Many companies and public services seem to be downsizing security to reduce spending. AI vision can play a vital role in boosting the effectiveness of smaller teams and solving issues like operator fatigue. The SOM-COMe-BT6-RK3588 offers a cost-effective solution for bringing advanced, power-efficient inference to local security camera systems, achieved through a highly integrated combination including the Metis AIPU, Voyager SDK, and Clea OS.
While security applications present a strong case for assisting operators with AI-enabled cameras, these solutions are also well-suited for factory automation and health-and-safety-conscious settings. Here, facial recognition, Re-ID, object detection, and body movement tracking can ensure workers are properly protected and remain fit to operate equipment for which they are qualified.
Learn more about SECO’s edge computing and CoM platforms at seco.com and discover how AI vision can be used on the SECO App Hub.