NVIDIA Jetson Orin NX Forklift & Pedestrian Monitoring Development Services

The First Engineering Task: Understand the Site, Not Just the Camera

A warehouse cannot be seen as a controlled environment. Warehouse camera performance can differ because of rack shadows, reflective floor surfaces, brightly colored clothing, the change of daylight at loading doors, suspended particles in the air, product movements and obstructions in people's footage that occur while filming in a warehouse.

Before developing the AI workflow, Adequate Infosoft would conduct a site assessment with the operations and safety teams. The assessment defines where the risk is, which behaviours the system should observe and what action is realistic when a warning occurs.

Questions that must be answered are as follows:

• How does the path of forklift and pedestrians cross or coincide?
• What intersections, corridors, or loading areas have no visibility?
• What are the typical speeds of the forklift and expected directions?
• Should there usually be pedestrians in the monitored area or it must remain empty?
• Can the property set up a permanent camera, lights, historical or limited access?
• How does the operator perceive a warning without losing attention?
• Who and what organizing actions is required to do after getting a record?

The camera location depends on the responses. For example, a camera mounted in the ceiling can show the general picture, but the details may not allow the identification of partly covered people.

A lower side camera may see a crossing more clearly but can be obstructed by pallets. In many sites, two complementary views are more reliable than one.

Each camera position is evaluated for field of view, mounting stability, lens selection, illumination, cable route, cleaning access and expected occlusion. For outdoor or loading-bay use, the enclosure, weather rating and low-light performance must be selected accordingly.

Edge Video Analytics with Jetson Orin NX

The Jetson Orin NX system receives one or more video streams from IP cameras, USB cameras or other supported industrial camera interfaces. The edge application performs detection locally and can operate without sending continuous video to the cloud.

The primary processing steps consist of:

The software can make use of NVIDIA TensorRT to optimize the detection model that was already approved. NVIDIA DeepStream can also be used in case of handling various streams of video, tracking and event metadata.

The final design depends on the number of cameras, required speed of processing, storage policy, and current state of the factory's IT infrastructure.

Proximity Is More Than an Image-Space Distance

A major design risk in camera-based proximity systems is assuming that two objects that appear close together in an image are physically close. Perspective distortion makes that unreliable without proper calibration.

For a fixed camera view, the system can be calibrated against known floor-plane reference points.

This allows pixel locations near the ground contact area to be mapped approximately into real-world coordinates. In more demanding deployments, stereo cameras, depth sensors, LiDAR, UWB or vehicle telemetry can be evaluated as complementary technologies.

The proximity logic should state what it actually measures. It may determine that a pedestrian and forklift occupy the same marked zone, that their projected paths are converging, or that they are within a calibrated approximate distance threshold.

It should not claim centimetre-level distance accuracy unless the deployed sensing method has been tested and verified for that condition.

The technology should not proclaim that the distance accuracy is measured down to the centimetre level until it is tested in that manner.

Good rules may be for example when:

• A person enters the active area for forklifts
• A forklift goes to the crossing when a person is present
• A person stays in restricted areas for too long
• Forklift reverses in the occupied area behind it
• Many people are present at the spot where a vehicle is moving
• The vision of the camera is blocked or is not fresh anymore

Alert Logic That Does Not Create Warning Fatigue

The system's alert design is as important as its AI model. Excessive alarms can distract operators, reduce trust and lead to alarms being ignored. Alerts should be meaningful, short, local to the risk area and tied to a response people understand.

At a fixed intersection, a stack light, directional warning lamp or audible beacon may be more useful than a remote push notification. A supervisor dashboard can provide incident records and recurring-risk analysis, but it is not a substitute for a real-time local warning.

The alert workflow has several phases:

In addition, if allowed by the factory conditions, the system can collaborate with PLCs, barricades, and speed regulators.

Thus, any device control action must be engineered separately and validated in terms of risks together with the equipment manufacturer and safety department of the plant.

Validation of the Model in Real Conditions

A model which has been trained on public images or clean sample datasets is not adequate for warehouse deployment, but rather specific videos collected directly from an actual site must be used for evaluation.

Our validation technique requires using clips from different shifts, uniforms of workers, types of forklifts, positions of pallets, lighting conditions, aisle congestion, and views of cameras.

At E-E-A-T, the difficult circumstances being tested include partial obstruction, pedestrians walking in reverse, reflective surfaces, worker carrying cargo and forklift operation with light or heavy load.

The project team produces a list of quantitative parameters that must be met at the time of testing including:

• Detection recall for people and lift trucks in the frame of reference of the camera
• Documented false-alarm rate per camera and per hour of operation
• Documentation of missed events
• Total delay in issuing alerts
• Percentage of recorded events
• Detection time for faults in camera or network
• Performance at fixed conditions of lighting and traffic
• Effectiveness of models after any modifications in layout or camera position

No machine vision system can be qualified as "collision-proof." A company practicing E-E-A-T expects its content on industrial safety to have descriptions that clearly explain system limits, hypothesis testing, and accountability of individuals.

Data Security, Privacy and Operational Data

Video analytics can involve personally identifiable information. When deploying video cameras, the system should state how it works, who can access it, how long is the information kept, what happens to the recorded event and how the data will be accessed.

Jetson devices are supposed to operate under secure credentials, have the software updated regularly and operate in a secure network environment.

The design determines how the IT department and the security guys may know everyone having access and how the solution works when connected to the Internet.