14.3.1 Two artificial-intelligence modes in Ergoniza Intelligent
Ergoniza Intelligent implements posture detection using two complementary approaches that address different privacy, speed, and accuracy requirements. The first, called Fast AI, performs inference on the user's device using code that runs in the browser. This approach considerably reduces response latency and, most importantly, avoids sending images and videos to Ergonautas servers, preserving the privacy of the audiovisual material at its source.
The second, called Accurate AI, delegates processing to Ergonautas server infrastructure. This option prioritizes estimation quality and robustness in complex scenarios, in exchange for longer processing time and the need to transfer data outside the local device. The user experience is consistent in both cases: the user provides the capture, reviews the result projected in the Ergoniza Intelligent interface, makes any appropriate adjustments, and immediately applies the detected posture data to the available assessment methods.
14.3.1.1 Fast AI (on the user's device)
In Fast AI mode, all detection is performed locally and media files never leave the user's device. This technical choice provides quick responses that facilitate iteration during capture—for example, repeating a shot or refining the framing—and prevents the selected image or video from being sent to Ergonautas servers for inference.
This mode supports both images and video sequences. For images, estimation focuses on a single posture representative of the captured moment. It immediately overlays anatomical landmarks and segments on the photograph and makes the resulting joint angles available in the interface. This facilitates a quick visual check and, if necessary, a specific correction by the evaluator.
For video, the system processes the complete sequence and maintains sufficient basic consistency between frames to observe the temporal evolution of a task, identify posture changes, and extract the metrics required for ergonomic analysis.
Practical performance depends on the device's capabilities—including its CPU and the acceleration available in the browser—the quality of the capture, and the absence of substantial occlusions. When lighting is uniform, the full body is visible in the frame, and a three-quarter view is used, local detection can provide suitable results for an initial estimate. The evaluator must review that estimate before making decisions, especially when speed and privacy are priorities.
14.3.1.2 Accurate AI (on Ergonautas servers)
Accurate AI mode transfers inference to Ergonautas servers, where specialized pipelines and computing resources can improve estimation quality under demanding conditions. In this approach, the user authorizes the image or video to be sent to the platform, which returns a more stable result and a more robust interpretation of complex scenes involving occlusions, less favorable viewpoints, or movements with extreme ranges.
For images, the system reconstructs the posture from a single photograph. This is especially useful when no video is available or temporal continuity is not required, and it provides broad anatomical coverage suitable for technical documentation.
For video, the platform generates a coherent time series of postures and angles, stabilized against spurious variations between frames. This facilitates the identification of work-cycle phases, exertion peaks, and limiting postures. Results are delivered in structured formats for direct integration into Ergoniza Intelligent's ergonomic assessment methods and for traceability in reports. The improved accuracy and robustness come at the cost of greater latency associated with data transfer and remote processing, which is generally acceptable when the primary objective is to obtain a higher degree of confidence in the estimate.
14.3.1.3 Image and video capture modes
In both Fast AI and Accurate AI modes, Ergoniza Intelligent can work with still images and video, adapting to the ergonomic problem being addressed. With images, the tool is designed to assess individual postures from representative photographs. The professional can clearly review the estimated overlay on the capture, confirm that key points are correct, and make adjustments if any misalignment due to occlusion or an unfavorable viewpoint is detected. This approach is particularly efficient for preparing technical records, documenting individual cases, and extracting static metrics without requiring temporal continuity. With video, the tool analyzes dynamic tasks and generates a posture sequence over time, making it possible to study kinematic evolution, detect critical moments, and compare the effects of changes in work methods or workstation design. The temporal consistency obtained in this mode facilitates the interpretation of movement patterns and provides richer context for risk assessment.
14.3.1.4 Workflow and selection criteria
The workflow is designed to be simple and consistent. First, the user selects the processing approach that best matches the priorities of the case: speed and privacy with local processing, or enhanced accuracy with server processing. The user then decides whether the case requires an image or video capture. After providing the material, the user views the estimated overlay returned by the system, confirms or corrects the relevant details, and immediately applies the posture data to the available assessment methods. The choice between the two approaches generally depends on the balance among response time, data sensitivity, and case complexity. Fast AI is appropriate when confidentiality and immediate results are the priorities. Accurate AI can be selected when remote processing is preferred and greater latency and transfer of the audiovisual material are acceptable. In both cases, the resulting posture must be reviewed and validated.
14.3.1.5 Capture best practices
Result quality depends substantially on how the capture is made. It is advisable to frame the person's entire body from head to toe, maintain uniform lighting that avoids backlighting and harsh shadows, and choose a viewpoint that reduces overlap and makes the relevant joints visible; a three-quarter view may be appropriate for many tasks. For video, the camera should be kept stable and the frame rate should be sufficient to describe the movement clearly. Following these guidelines increases estimation reliability in both Fast AI and Accurate AI modes and reduces the need for subsequent manual corrections.