This project implements a modular and scalable Computer Vision (CV) system designed to replace traditional physical barriers, enhancing worker safety in industrial settings (Industry 4.0/5.0). The core objective is to achieve contextual control of machinery based on the dynamic state of the surrounding work environment.

The system continuously monitors a defined workspace. Upon the detection of personnel entering this area, appropriate countermeasures are instantly triggered, influencing machinery behavior to prevent hazardous situations. This compartmentalized architecture promotes high modularity and scalability.

The architecture is built on flexible components, ensuring low-latency processing critical for safety applications.

For flexible and high-performance monitoring, the system utilizes Industrial Cameras.

• Flexibility & Performance: The camera can be positioned dynamically within the workspace while providing the necessary data quality for vision processing.
• Protocol: Data is transmitted over a standard network using the GigE Vision protocol to ensure stable and responsive data streaming.

Effective workspace monitoring requires sophisticated data processing. This computation is realized on powerful hardware situated on-site, leveraging the Edge Computing paradigm to guarantee the low-latency required for real-time safety.

• Human Recognition: A fast and efficient ONNX object detection model (YOLO11) identifies the presence of personnel.
• Contextual Awareness: High-performance ONNX depth estimation models (UniDepth v2 or DepthAnything v2) are used to calculate distances, enabling more granular control responses.
• Hardware & Portability: By converting models to the ONNX runtime format, is ensured hardware-agnostic deployment. The system is optimized to achieve the best performance with the available hardware and optimisations.

Based on the computed results (detection and distance), a core logic module interprets the data and transmits new directives to the industrial machinery.

• Use Case: The system manages the speed of a Universal Robots robotic arm by sending new velocity values (using the RTDE protocol) to dynamically increase, decrease, or stop the robot, ensuring safety based on the worker's position.

The architecture operates under the Edge Continuum principle to guarantee real-time performance.

The overall system architecture is conceptually represented below, illustrating the relationship between sensing, processing, and control elements.

Fig. 1 - General System Architecture.

The functional flow illustrates the complete, contextually driven sequence from data acquisition to the final control action:

Fig. 2 - Functional Flow of the Architecture.

For future scalability or in highly distributed environments requiring deterministic timing guarantees, the system is prepared for integration with the Time-Sensitive Networking (TSN) protocol.

• TSN Benefit: The implementation of TSN (requiring a specialized switch and camera support) would provide deterministic latency and prioritize data flow, further strengthening the system's real-time performance for safety-critical applications.

This repository serves as the super-repository for the EdgeCV4Safety project, utilizing Git Submodules to maintain strict modularity between the processing and control components.

The complete system is divided into two distinct, independently managed components:

Fig. 3 - Project Submodules.

The specific documentation for each submodule is available on respective pages.
Update: now repos are private (due to law/contracts), but you can acces anyway to each README (migrated here, see table).

To clone the main repository and automatically download the contents of both submodules, use the --recurse-submodules flag:

git clone --recurse-submodules https://github.com/justwhiteecode/EdgeCV4Safety.git
cd EdgeCV4Safety

If you have already cloned the repository without the --recurse-submodules flag, you can initialize and download the submodules manually:

git submodule update --init --recursive

After cloning, please refer to the specific setup instructions within each submodule directory for dependency installation:

• Vision System Setup: See the README.md and requirements.txt in the vision_processing/ directory.

• Controller System Setup: See the README.md and requirements.txt in the robot_controller/ directory.