Hi team, I have a couple of embedded CV related ideas to prototype. I almost bought a sperate camera + raspberry pi kit and saw this product on ArduiCam. It looks super exciting! Some questions on my idea feasibility & what are the best components to buy for prototyping? 1. Object detection on glasses for vision impaired pedestrian. (Mainly target China market). - OAK-D seems the best fit. However, the processing board is attached to the camera modules and I don't know if it's going to be too large? Ideally I want cameras to face front but boards on the side so the speakers can be next to the ears. Thoughts on if I can buy modular camera model with the kit? - I saw OCR was included. I'm assuming I can just load any chinese OCRs for the target audience? 2. Detecting the onset of epilepsy in a medical settings. Mostly I want to zoom in to patients face when it happens and potentially create alerts for the healthcare providers. - This seems better suited to get OAK-1? - How wide is the camera angle? Lastly, are there any power metrics (how much battery does it need for a 4hr run)? I originally was thinking just using pi to stream video to mobile phones / PCs and off load ML models there. Have y'all done some comparison on both power and latency? Thanks! Mira

Hello Mira! Thanks for the kind words:) Regarding the object detection, you could also use the [DepthAI FCC](https://docs.luxonis.com/en/latest/pages/products/bw1098ffc/) module, to which you can attach cameras. It was designed specifically for such use cases. Note here that you will have to calibrate the mono cameras for the depth ([some info here](https://docs.luxonis.com/en/latest/pages/faq/#id3)), so a bit more work. And you should also check the `Vision System for visually impaired` from the [OpenCV competition](https://opencv.org/opencv-spatial-al-competition-winners-announced/) for some ideas. We do have an [OCR demo](https://github.com/luxonis/depthai-experiments/tree/master/gen2-ocr) on the `depthai-experiments` repo, and as you said, it could be modified for your use-case. I'm sure there are some pre-trained models for Chinese OCR. For the epilepsy detection, [OAK-1](https://docs.luxonis.com/en/latest/pages/products/bw1093/) should do the job, since you don't need the spatial information of the patient. Note here that there is no "mechanical zoom " on the camera, so to get a high enough resolution of the face I would first do a facial detection on preview image (for example 300x300 and feed the preview to the pretrained `face-detection-retail-0004` model, similar [demo here](https://github.com/luxonis/depthai-experiments/blob/gen2-triangulation-demo/gen2-triangulation/main.py)) and then based on the face detection result take the 4k video and crop that to get a high-resolution image of the detected face - to feed it into your epilepsy detection model. And as per the OAK-1 docs, the FoV is `81° DFoV - 68.8° HFoV`. Power consumption mostly depends on the workload of the Myriad X (the VPU), but from our testing OAK-1 uses about `~0.6A @ 5V`. Thanks, Erik

Thoughts on which kits to buy for application prototyping?

lucy23shirley

Hi team,

I have a couple of embedded CV related ideas to prototype. I almost bought a sperate camera + raspberry pi kit and saw this product on ArduiCam. It looks super exciting! Some questions on my idea feasibility & what are the best components to buy for prototyping?

Object detection on glasses for vision impaired pedestrian. (Mainly target China market).
- OAK-D seems the best fit. However, the processing board is attached to the camera modules and I don't know if it's going to be too large? Ideally I want cameras to face front but boards on the side so the speakers can be next to the ears. Thoughts on if I can buy modular camera model with the kit?
- I saw OCR was included. I'm assuming I can just load any chinese OCRs for the target audience?
Detecting the onset of epilepsy in a medical settings. Mostly I want to zoom in to patients face when it happens and potentially create alerts for the healthcare providers.
- This seems better suited to get OAK-1?
- How wide is the camera angle?

Lastly, are there any power metrics (how much battery does it need for a 4hr run)? I originally was thinking just using pi to stream video to mobile phones / PCs and off load ML models there. Have y'all done some comparison on both power and latency?

Thanks!
Mira

erik

Hello Mira!
Thanks for the kind words🙂 Regarding the object detection, you could also use the DepthAI FCC module, to which you can attach cameras. It was designed specifically for such use cases. Note here that you will have to calibrate the mono cameras for the depth (some info here), so a bit more work. And you should also check the Vision System for visually impaired from the OpenCV competition for some ideas. We do have an OCR demo on the depthai-experiments repo, and as you said, it could be modified for your use-case. I'm sure there are some pre-trained models for Chinese OCR.

For the epilepsy detection, OAK-1 should do the job, since you don't need the spatial information of the patient. Note here that there is no "mechanical zoom " on the camera, so to get a high enough resolution of the face I would first do a facial detection on preview image (for example 300x300 and feed the preview to the pretrained face-detection-retail-0004 model, similar demo here) and then based on the face detection result take the 4k video and crop that to get a high-resolution image of the detected face - to feed it into your epilepsy detection model. And as per the OAK-1 docs, the FoV is 81° DFoV - 68.8° HFoV.

Power consumption mostly depends on the workload of the Myriad X (the VPU), but from our testing OAK-1 uses about ~0.6A @ 5V.

Thanks, Erik

lucy23shirley

Awesome! Thanks for the quick response Erik and all the resources you linked! Will get the kits and excited to report back afterwards!! 😀