Edge AI Models

Traditional sensor systems often reach their limits when it comes to reliable person detection: Rule-based methods have limited capabilities for detecting people, while camera-based solutions are often sensitive to changing lighting conditions, obstructions, or complex environments.

This AI model for person detection identifies people in real time directly on edge hardware. It performs reliably under varying lighting conditions, movements in the environment, and different spatial structures.

Fully local processing enables extremely short response times, protects sensitive data, and allows for flexible integration into a wide variety of systems to ensure reliable detection and tracking of people across many industries.

Technical Specifications:

Person Detection on Aarch64:

• RAM Required: 838.46 kB
• ROM Required: 1250.64 kB
• Inference Time: 14 ms

Person detection on Cortex-M7:

• RAM required: 387.96 kB
• ROM required: 474.45 kB
• Inference time: 480 ms

Person detection on x86:

• The model has not been optimized for execution on x86 computers and is intended for use on this architecture solely to test the model’s functionality if the embedded hardware is not yet available.

Features:

• Energy efficiency: Up to a thousand times lower power consumption than GPU solutions
• Local processing: Millisecond-fast responses without cloud latency
• Data protection through Edge AI: All image data remains on-site → no external transmission
• Network independence: Works even without a stable internet or cloud connection

Keyword Spotting enables the reliable detection of predefined keywords or short voice commands in audio signals—continuously, in real time, and in a resource-efficient manner.

The AI model runs directly on edge hardware and recognizes relevant keywords even in noisy environments, with different speakers, and varying pronunciations.

Local processing enables extremely short response times, while sensitive audio data never leaves the system. The model is designed for low energy consumption and is suitable for continuous operation in embedded systems, mobile platforms, and autonomous machines.

Features:

• Energy efficiency: Up to a thousand times lower power consumption than GPU solutions
• Local processing: Millisecond-fast responses without cloud latency
• Data protection through Edge AI: All data remains on-site → no external transmission
• Network independence: Works even without a stable internet or cloud connection

In many industrial facilities, key quality assurance processes are still carried out manually—a time-consuming, error-prone process that relies heavily on skilled personnel. At the same time, demands are rising: greater precision, lower scrap rates, and increasing pressure to improve efficiency. Modern AI can help here by automatically analyzing components and detecting anomalies early and reliably.

The AI model for anomaly detection identifies deviations from a known normal state in image or sensor data—reliably, early, and without explicit error classification.

Unlike traditional inspection methods, the model is trained exclusively on error-free examples. This eliminates the need for extensive defect datasets or labor-intensive labeling processes, and even previously unknown defect patterns can be reliably identified.

Technical Specifications:

Anomaly detection on Aarch64:

• Required RAM: 5767.17 kB
• Required ROM: 4082.54 kB
• Inference time: 430 ms

Anomaly detection on x86:

The model has not been optimized for execution on x86 computers and is intended for use on this architecture solely to test the model’s functionality if the embedded hardware is not yet available.

Features:

• Energy efficiency: Up to a thousand times lower power consumption than GPU solutions
• Local processing: Millisecond-fast responses without cloud latency
• Data protection through Edge AI: All image data remains on-site → no external transmission
• Network independence: Works even without a stable internet or cloud connection