OAK 4 D PRO night vision image quality

JanMartnek

Hello,

We are designing a system for hospital patient monitoring, and the OAK-4D Pro camera seems like a promising hardware choice for our specific use case. However, I am particularly concerned about its image quality in low-light or near-dark conditions, as the system needs to operate reliably at night.

Would it be possible to access somewhere a full-resolution sample image taken under such conditions? It’s important for us to confirm that sufficient facial details can be captured from a distance of approximately 2–3 meters.

Many thanks!

BR,
Jan

jakaskerl

JanMartnek
Would this work?

Is it possible to use OV9282 + a floodlight/high exposure time?
The IMX586 low light peformance isn't the best, but we could swap out the main color sensor for a STARVIS sensor or something that would allow better low-light performance.

Thanks,
Jaka

JanMartnek

jakaskerl

Thank you for your reply!

Unfortunately, using the OV9282 sensor with a long exposure time doesn’t seem suitable for our use case. I need to highlight an important requirement: detecting the blink of an eye in addition to other facial characteristics. While fast eyelid movements or transitions are not critical, we must accurately distinguish between open and closed eyes and measure the duration of each state with a resolution of at least 0.1 seconds. Therefore, a video capture frame rate of 30 fps or higher would be ideal.

For reference, we are currently experimenting with the OAK-D Pro, utilizing its integrated IR LED illumination. As shown in the attached picture, the image quality is quite poor, and the eyelid state cannot be reliably recognized by inference with our neural network. Even though the attached image is resized, the full-resolution image doesn’t offer significant improvement. We believe the unreliable inference is primarily due to the poor image quality.

Based on our experience with the OAK-D Pro, we are inclined to think that the OAK-4D Pro could deliver better image quality, in addition to offering broader possibilities for running neural networks.

We would greatly appreciate any suggestions or improvements (such as using a STARVIS sensor) that could enhance image quality.

jakaskerl

JanMartnek Unfortunately, using the OV9282 sensor with a long exposure time doesn’t seem suitable for our use case. I need to highlight an important requirement: detecting the blink of an eye in addition to other facial characteristics. While fast eyelid movements or transitions are not critical, we must accurately distinguish between open and closed eyes and measure the duration of each state with a resolution of at least 0.1 seconds. Therefore, a video capture frame rate of 30 fps or higher would be ideal.

The images in the link I sent you, set the OV9282 exposure to 30ms which should be fine even for 30FPS.
Would it make sense to get a more powerful IR light (940nm) to use instead of the IR floodlight on the device which is quite weak? Or would that affect the patient too much?

JanMartnek We believe the unreliable inference is primarily due to the poor image quality.

That makes sense.. I'd still check the ISO and lower it to reduce noise which interferes with detection accuracy.

JanMartnek Based on our experience with the OAK-D Pro, we are inclined to think that the OAK-4D Pro could deliver better image quality, in addition to offering broader possibilities for running neural networks.

OAK4-D pro still features the same OV9282 mono sensors for depth imaging. It's however possible to switch the center RGB for a IMX462 or some other StarVis sensor, if enough devices are bought. See docs.

Best to test with a FFC-3/4P + IMX462 (or OAK-T which also has the IMX462) if the quality is good enough for detection.

LMK your thoughts.

Thanks,
Jaka

JanMartnek

Thank you for your response.

We have already attempted to boost 940nm illumination using two 1W LEDs with the current OAK-D Pro (OV9282) setup, but unfortunately, we didn’t see any significant improvement in image quality. This level of illumination power shouldn’t mean any risk to the patient, though further testing and qualification would be required. For now, our intention was simply to evaluate the immediate effect of the additional IR light source.

I'll check with our SW guy which ISO value was used setting just to make sure that it's not set unreasonably high.

That said, experimenting with the IMX462 sounds exciting. If I understand correctly, it’s possible to replace the center sensor of the OAK-D Pro with the IMX462 in a DIY manner, which would be the quickest way to get initial results. If this is feasible, could you please recommend a suitable CCM with the IMX462 that would fit into the OAK-D Pro housing, assuming such an option exists?

Many thanks!

Jan

jakaskerl

JanMartnek If I understand correctly, it’s possible to replace the center sensor of the OAK-D Pro with the IMX462

We only have access to M12 IMX462 ccms, which won't fit inside the D-pro case. Best to ask Arducam for a non-M12 version. Easiest ways would be a FFC modular device or try with OAK-T.

Thanks,
Jaka