AI on the edge for vision systems

Questions?

Need samples?

Send us a sample request

Worth reading

RAFI FS series modular control & signal system

Application Specific Batteries

RRC Power Management Modules

In-depth

AI vision enables systems to analyse and interpret visual data to support automated decision-making. From industrial inspection to robotics and retail analytics, AI vision enhances performance, efficiency and system autonomy, allowing systems to interpret and act on visual data with greater consistency and speed compared to traditional rule-based approaches.

When deployed on the edge, AI vision systems process data locally, enabling real-time decision-making without reliance on cloud infrastructure.

Development teams often face challenges in selecting the right hardware, integrating AI models and achieving reliable performance within system constraints. In edge deployment, this becomes even more complex, as development teams must identify suitable hardware that meets the required performance, while balancing compute capabilities, power consumption and system constraints. Texim supports OEMs in selecting, optimising and integrating embedded AI platforms tailored to their application requirements.

AI vision creates value in applications for various markets, where visual data must be processed reliably, consistently and in real time.

Request AI vision integration

AI inference at the edge enables real-time visual inspection and process monitoring directly within industrial equipment. By integrating AI-capable embedded systems, manufacturers can analyse high-resolution image streams without relying on external compute infrastructure. This improves inspection performance and system efficiency, including:

Real-time defect detection
Inspection directly on the edge
Reduced bandwidth usage
Energy-efficient processing

Ai On Edge, Industrial, 16x9

Ai On Edge, Medical, 16x9

By enabling rapid preliminary image analysis within the device, AI vision can support earlier pattern identification, accelerate diagnostics and generally reduce operational pressure on medical personnel.

In medical systems, edge acceleration enables OEMs to integrate advanced image processing directly into their equipment:

Faster preliminary image analysis
Improved workflow efficiency
Consistent image evaluation
Secure on-device processing

Vision AI enables automated monitoring and analysis in high-throughput commercial environments. Edge acceleration ensures that processing remains fast and independent of network conditions:

Real-time object recognition
No-cloud video analytics
Lower bandwith requirements
Reliable facility operation

Ai On Edge, Retail, 16x9

Ai On Edge, Robotics, 16x9

Robotics platforms increasingly rely on computer vision workloads that require high-performance, energy-efficient inference at the edge. By integrating AI-capable embedded systems, visual data can be processed locally with low latency and without continuous cloud dependency. This approach provides several advantages for robotic platforms:

Environment awareness
On-device image processing
Scalable deployment
Optimised performance

AI vision systems are typically designed as a combination of hardware, software and AI processing components. This section outlines how these systems are structured in practice, from core building blocks to implementation considerations.

AI vision systems consist of multiple functional components, including image acquisition, processing, AI inference and system integration.

These building blocks determine how data flows through the system and how performance requirements are met.

In practical implementations, these components are distributed across the system architecture, where general-purpose processors handle control and pre- and post-processing tasks, while dedicated AI accelerators perform inference. This structured approach forms the basis for how AI workloads are organised in pipelines on edge devices.

AI vision systems often rely on structured processing pipelines to handle multiple tasks efficiently. These workloads are typically implemented on edge devices as multi-stage pipelines. Depending on application requirements, they can be organised as cascaded or parallel processing flows.

In cascaded pipelines, the output of one model is used as input for the next stage. For example, object detection can trigger follow-up tasks such as text recognition, classification or anomaly detection. In parallel pipelines, multiple models process the same input at the same time, combining insights from different vision tasks.

This approach enables efficient use of compute resources while supporting predictable, real-time processing within embedded systems.

AI workloads are typically distributed across the main processor and a dedicated AI accelerator.

Pre- and post-processing of image and video data is handled by the main CPU, while the Metis AIPU acts as a co-processor for AI inference. This architecture enables efficient task distribution between general-purpose processing and AI acceleration, supporting low-latency AI vision deployments.

This diagram illustrates how the Axelera Metis AI Processing Unit (AIPU) functions within such a system.

Axelera Metis

Ai On Edge Implementation Flow

Designing an AI vision system involves multiple steps, including model selection and training, system integration and deployment on edge hardware. These stages are closely linked, and decisions made early in the process directly impact performance, reliability and scalability in real-world applications.

This diagram shows a flow that focuses on the implementation phase, but this represents only one part of the overall system design. Within the broader design process, multiple approaches and architectural choices are possible depending on application requirements.

Model choice and training

AI vision performance depends on training data quality and model development, including deployment. A well-trained model determines whether a system detects subtle defects, adapts to changing environments or struggles with false positives.

Effective AI vision systems do not start with hardware; they start with representative data and structured model development. The reliability and accuracy of the final application depend directly on how the model is trained, validated and maintained over time.

In practice, effective training requires:

Representative, high-quality image datasets
Correct labelling and data preparation
Selection of appropriate model architectures
Continuous validation and retraining when conditions change

For industrial applications, this is particularly critical. Lighting variations, product tolerances and environmental factors directly affect model accuracy. Without proper training and optimisation, even advanced hardware cannot deliver reliable results.

Integration

Integrating AI vision models into embedded systems requires alignment between hardware capabilities, software frameworks and real-time processing requirements. This includes selecting suitable processors, managing data flows and ensuring reliable performance within system constraints.

Deployment and optimisation

Deploying AI vision systems involves translating validated models into real-world applications by aligning hardware, software and operational requirements, including model conversion, optimisation and runtime execution on edge hardware.

The Axelera AI Voyager SDK provides the software environment required to develop, deploy and optimise AI vision applications on edge devices. It supports both rapid prototyping and production deployment, enabling teams to work with pre-trained models or integrate and adapt their own models, while optimising performance for specific edge hardware.

The SDK includes access to a Model Zoo with validated models that can be fine-tuned for specific use cases. It also supports runtime development, allowing existing models to be extended or optimised based on application requirements. This enables teams to move quickly from evaluation to a working system using validated, ready-to-use building blocks.

Axelera Voyager

AI vision systems can be configured using different hardware architectures. In practice, the optimal setup depends on performance requirements, power constraints and available system interfaces. These configurations reflect common approaches to deploying AI vision at the edge across different performance and integration requirements.

N010 Inside

N010 Inside 2

Lenovo ThinkStation P360 ULTRA with a Metis PCIe

x86-based edge AI systems with PCIe acceleration

For high-performance AI vision applications such as industrial inspection, robotics or multi-camera systems, x86-based platforms combined with PCIe AI accelerators are a commonly used approach. These systems are designed to provide the compute capacity to support complex models and high-throughput data processing requirements in AI vision applications.

In this architecture, the main CPU handles system control as well as pre- and post-processing tasks, while a dedicated AI accelerator performs inference. This separation enables efficient workload distribution and is designed to support low-latency processing in demanding environments.

Typical system components include:

x86-based box PC or industrial workstation
PCIe AI-accelerator (AIPU)
Camera or sensor input
Local edge processing without reliance on cloud infrastructure

Example configurations include systems such as the Dell Precision 3460XE and industrial platforms like the Advantech ARK-3534B, each combined with a Metis PCIe AIPU to support dedicated AI inference acceleration alongside general-purpose processing.

Compact edge AI systems with M.2 acceleration

For space-constrained or power-sensitive deployments, compact edge systems using M.2-based AI accelerators offer an alternative. These configurations are commonly used in embedded applications where size, integration and energy efficiency are critical.

In this setup, the embedded platform manages system control and data handling, while the M.2 accelerator enables on-device AI inference. This supports real-time processing close to the data source on the device itself.

Typical system components include:

Embedded system or compact industrial platform
M.2 AI accelerator (AIPU)
Integrated camera or imaging interface
Fully local processing at the edge

Example implementations include solutions such as the Metis M.2 1x AIPU, 1 GB LPDDR4x, enabling on-device AI inference within compact system designs.

Integrated AI compute boards

In integrated environments, compute boards with onboard processing and memory can function as standalone AI vision platforms. These systems are suitable for applications where a compact, self-contained solution is preferred.

Such architectures combine general-purpose processing and AI capabilities within a single board, enabling simplified system design and faster integration. While typically offering lower peak performance than PCIe-based architectures at the edge.

Typical system components include:

Integrated compute board (e.g. ARM-based platform)
Onboard CPU and memory
Camera interfaces and I/O connectivity
Edge-based processing within a single device

These configurations are often used in entry-level or space-limited applications where moderate AI performance is sufficient.

Example implementations include platforms such as the Metis SBC with RK3588 and onboard AIPU (Mini-ITX), which can be used as a compact and self-contained platform for deploying AI vision applications at the edge.

E010 Top

D010 Top

Other examples of Metis-accelerated systems (Dell PRECISION 3460XE and Advantech ARK-3534B)

Selecting the appropriate system architecture depends on factors such as model complexity, latency requirements, power consumption and form factor constraints. In practice, balancing these aspects is essential to achieving reliable and scalable AI vision performance at the edge.

Our engineering team works with customers throughout the AI vision system design, integration and optimisation process. We collaborate closely with development teams to ensure the final solution aligns with defined performance targets and product requirements.

We focus on enabling development teams rather than delivering full application projects. This includes providing the software stack, tools and technical know-how required to get AI vision systems running efficiently on edge hardware, supporting initial evaluation and enabling further independent development.

Our approach helps reduce complexity and shorten development time while allowing teams to retain control over their own implementation. Read more about how we approach integration and ongoing support.

– This page was reviewed by Mels Liefaard, a technical B2B specialist with over 10 years of experience in industrial applications and embedded systems.

What is AI vision?

AI vision is the application of artificial intelligence, typically deep learning models, to process and interpret visual data from cameras or imaging systems. It enables machines to detect objects, identify patterns and extract meaningful information in real time, either on embedded edge hardware or in cloud infrastructure.

In practice, the workflow typically begins with capturing visual input via cameras or other sensing devices. The collected data is then prepared through preprocessing steps, such as resizing, normalisation or basic enhancement. A trained deep learning model subsequently analyses the processed input and performs inference, producing outputs such as classifications, detected objects, extracted text or identified activities.

Depending on requirements such as latency, bandwidth and data handling, the solution can be deployed and operated either on edge hardware or within cloud infrastructure.

What is AI on the edge?

This technique refers to running AI models locally on devices, gateways or edge infrastructure instead of sending all data to the cloud for processing.

Why does AI on the edge matter for B2B organisations?

AI on the edge entails several advantagse:

Lower latency: Enables near real-time decisions in operational environments
Reduced bandwith and cloud costs: Only relevant results are sent to the cloud
Improved data control: Sensitive data can remain on-site

Edge AI is particularly relevant where speed, reliability and data controle are requirements.

How do you optimize AI models for edge devices?

Optimising AI models for edge deployment requires balancing accuracy, memory usage and power consumption. Unlike cloud environments, embedded platforms operate under strict processing and thermal constraints.

This typically includes selecting model architectures suited for resource-constrained environments, applying specific techniques to help decrease memory usage and power consumption while maintaining acceptable accuracy. Performance validation on the target device is essential to ensure stable performance under real operating conditions.

What are the risks and limitations of AI on the edge?

While AI on the edge provides operational advantages, it also introduces specific technical challenges:

Hardware constraints: Edge devices often have limited computer power, memory and energy capacity, which restricts model size and complexity
Security exposure: A distributed device fleet expands the attack surface and required device authentication, encryption and update mechanisms
Higher upfront investments: Edge infrastructure and fleet management toolins may require additional capital expenditure compared to cloud-only approaches

Which data protection considerations apply to AI vision systems?

In applications where AI vision systems process images that may contain identifiable individuals, applicable data protection regulations, including GDPR, must be taken into account. Depending on context and use, visual data can qualify as personal data. Where this is the case, a lawful basis for processing and clearly defined purpose limitation must be established.

System developers are responsible for implementing appropriate data governance measures, including anonymisation, access controls and secure data handling procedures.

Edge-based inference can support privacy-oriented system architectures by minimising external data transmission and reducing reliance on external cloud storage.

In-depth

RAFI FS series modular control & signal system

The RAFIX 22FS+ and RAFIX 30FS+ is a modular two-part system for “wire to” fast on connection, or PCB’s. Designers can choose between the ultra-low profile PCB contact blocks with a 9.2mm back plane, or the wired connections, which can be soldered or accept 2.8mm x 0.5mm standard connector. For the HMI actuator or indicator […]

In-depth

Application Specific Batteries

VARTA’s Application Specific Batteries (ASB) are designed for small and medium-sized vehicles like AGV and robotics. Featuring modular and expandable Lithium-Ion batteries, they’re adaptable to various applications, offering a scalable range from 0.58kWh to 37.5kWh. The EasyPro series use Lithium Iron Phosphate (LiFePO4) for enhanced safety and long cycle life, while the EasyBlade and EasyBlade […]

In-depth

RRC Power Management Modules

Power Management Modules (PMMs) from RRC Power Solutions handle the switch between AC/DC adapter power and battery backup in case of input power failure. These compact, certified modules handle charging, SMBus communication and battery management. No firmware needed, and no certification work. A PMM makes the difference when: Your device charges a smart battery It auto-switches […]

In-depth

High-precision IMUs

Together with Epson, Texim Europe specializes in precise IMU solutions for industrial, robotics, industrial automation, and aero(space) systems. With our expertise in advanced QMEMS technology and IP67-rated protection, we deliver reliable performance tailored to even the toughest environments.

In-depth

Globally certified batteries

RRC power solutions smart battery packs RRC power solutions brings reliability and innovative technologies to mobile power supplies in the form of standard energy storage products. These batteries are engineered to meet the highest global regulatory standards, providing seamless integration for medical and industrial applications worldwide. With comprehensive certifications across Europe, North America, Asia-Pacific, and […]

In-depth

Connecting IoT devices made easy

Tired of complicated (remote) IIoT setups? This CaaS-platform (Connectivity-as-a-Service platform) is your answer. Choose whether you like a plastic triple-cut SIM, eSIM, or SoftSIM. SoftSIM is the most flexible option for future-proofing your devices. It eliminates the need for physical cards, optimizes battery life, and simplifies IoT deployments.

In-depth