Hello Luxonis team and community, We are currently developing an AI-based Proof of Concept (PoC) for an electrical company in Peru, South America, focused on automated detection of Minimum Safety Distance (DMS) risks associated with Medium Voltage (MV) electrical distribution networks in Lima, Peru. As part of this project, we have been extensively testing the OAK-D Pro W PoE - IMX378 under real-world operational conditions, including: * Vehicle-mounted outdoor capture * Urban environments with shadows and variable lighting * Continuous movement and vibration * Dense infrastructure scenarios with cables, poles, buildings, signs, etc. Our AI workflow is based on: * YOLO models * Roboflow training pipelines * Proprietary post-processing software We trained approximately 15 risk-related classes, including: * Scaffolding * Ladders * Buildings * Streetlights * Advertising signs * Forklifts * Dump trucks * Other urban risk elements near MV lines Additionally, we trained: * MV poles * Insulators * MV support structures Our approach consists of: 1. Detecting risk elements 2. Detecting MV infrastructure 3. Relating both detections spatially 4. Estimating distances using stereo depth + interpolation The object detection itself works surprisingly well. However, we are currently facing significant limitations regarding depth estimation accuracy in real outdoor deployments. ## Main observations from field tests ### 1. Long-range depth instability Although documentation mentions depth capabilities up to \~15m, in our real-world tests we are not obtaining stable or usable measurements beyond \~10–12m. In practice, when objects are located between 15m and 40m: * depth becomes unstable, * noisy, * or simply unusable for DMS evaluation. ### 2. Large real-world depth errors Even within closer ranges (

Hey @"PabloPerezLeiva_innspatial"#p33785 - thanks for the detailed write-up. For this use case, I would look at [OAK-D LR](https://docs.luxonis.com/hardware/products/OAK-D%20LR) much sooner than trying to push [OAK-D Pro W PoE](https://docs.luxonis.com/hardware/products/OAK-D%20Pro%20W%20PoE) further, because we explicitly recommend LR for long-range stereo in the [Configuring Stereo Depth](https://docs.luxonis.com/hardware/platform/depth/configuring-stereo-depth/) docs due to its **15 cm baseline** and narrower default optics. That said, even with [OAK-D LR](https://docs.luxonis.com/hardware/products/OAK-D%20LR), the published [Depth Accuracy](https://docs.luxonis.com/hardware/platform/depth/depth-accuracy/) guidance is still much closer to the `~7.5–12.5 m` band than to `15–40 m` safety-critical ranging. Luxonis shows roughly `

OAK-D Pro W PoE depth accuracy limitations Lima-Peru

PabloPerezLeiva_innspatial

Hello Luxonis team and community,

We are currently developing an AI-based Proof of Concept (PoC) for an electrical company in Peru, South America, focused on automated detection of Minimum Safety Distance (DMS) risks associated with Medium Voltage (MV) electrical distribution networks in Lima, Peru.

As part of this project, we have been extensively testing the OAK-D Pro W PoE - IMX378 under real-world operational conditions, including:

Vehicle-mounted outdoor capture
Urban environments with shadows and variable lighting
Continuous movement and vibration
Dense infrastructure scenarios with cables, poles, buildings, signs, etc.

Our AI workflow is based on:

YOLO models
Roboflow training pipelines
Proprietary post-processing software

We trained approximately 15 risk-related classes, including:

Scaffolding
Ladders
Buildings
Streetlights
Advertising signs
Forklifts
Dump trucks
Other urban risk elements near MV lines

Additionally, we trained:

MV poles
Insulators
MV support structures

Our approach consists of:

Detecting risk elements
Detecting MV infrastructure
Relating both detections spatially
Estimating distances using stereo depth + interpolation

The object detection itself works surprisingly well.

However, we are currently facing significant limitations regarding depth estimation accuracy in real outdoor deployments.

Main observations from field tests

1. Long-range depth instability

Although documentation mentions depth capabilities up to ~15m, in our real-world tests we are not obtaining stable or usable measurements beyond ~10–12m.

In practice, when objects are located between 15m and 40m:

depth becomes unstable,
noisy,
or simply unusable for DMS evaluation.

2. Large real-world depth errors

Even within closer ranges (<10–12m), the error margin remains high for our application.

Example:

System estimation: ~3m
Real measured distance: ~1.5–2m

For industrial electrical safety applications, this error is still too large.

Questions for the community / Luxonis team

Are these results expected for stereo depth under real outdoor urban conditions?
Has anyone successfully achieved sub-meter accuracy in:
moving vehicle deployments,
outdoor environments,
medium/long-range depth estimation?
Are there recommended:
stereo configurations,
calibration workflows,
post-processing filters,
depth presets,
or disparity settings
that significantly improve outdoor performance?
Would another Luxonis device perform better for this use case?
We are particularly interested in:
long-range outdoor depth,
moving platforms,
urban environments,
and higher spatial accuracy.
At this point, should we assume stereo depth alone is insufficient for this type of application and start integrating:
LiDAR,
SLAM,
RTK,
or multi-sensor fusion?

We would really appreciate any feedback, references, or recommendations from the community.

Here is a video showing some examples of the system operating in real conditions:
https://www.youtube.com/watch?v=A96s0WjHiQE

Thanks a lot.

Pablo - Innspatial team

OskarSonc

Hey PabloPerezLeiva_innspatial - thanks for the detailed write-up. For this use case, I would look at OAK-D LR much sooner than trying to push OAK-D Pro W PoE further, because we explicitly recommend LR for long-range stereo in the Configuring Stereo Depth docs due to its 15 cm baseline and narrower default optics.

That said, even with OAK-D LR, the published Depth Accuracy guidance is still much closer to the ~7.5–12.5 m band than to 15–40 m safety-critical ranging. Luxonis shows roughly <1.5% error below 7.5 m and <2.5% from 7.5–12.5 m for LR. The LR product page also includes a theoretical table showing that with the default 60° HFOV, <2% error is around 8.3 m and <4% error around 20.8 m, but that is not the same as saying it will deliver reliable, field-ready DMS measurements at those distances in a moving outdoor urban scene.

So for your current OAK-D Pro W PoE, the behavior you see is broadly expected: it is a wide-FOV stereo camera with a 75 mm baseline, and the published Depth Accuracy numbers for that family fall off much earlier. Also, as noted in ToF vs Stereo, active stereo helps much less outdoors because the IR dot pattern is not very useful in sunlight.

If you continue testing the current device, I would still try the recommendations from Configuring Stereo Depth: enable Subpixel and Left-Right Check, use an accuracy-oriented preset, and verify calibration health with Dynamic Calibration.

Thanks,
Oskar