OAK-D PoE camera working very slowly and very laggy

Fakhrullo

Hi guys, this is my first time working on the projects like this but I usually work on web development (backend).

So I got the gen2-cumulative-object-counting project from the official sources GitHub page and custom-sized it a lot. The main purpose of my project is to count how many people entered and went out of the room, and should send the collected data every 5 minutes. I implemented these functions everthing is working fine but the streaming is so laggy and because of that not counting correctly. I'm using an OAK-D PoE camera for my project. This project is for production purposes. Please help me in advance, I need your help guys. Here's the basic understanding of the project if a person passes both green lines it should count +1 or -1

So why it's working so slowly? is there any optimizations I should do or do some testing? Every thing is working fine but so slow

Fakhrullo

the image is just for demonstration, Actual position of the camera will be above the door, so the camera now is upside looking to the floor

jakaskerl

Hi Fakhrullo
Might it be the bandwidth requirements for the project that cause the low FPS?
https://github.com/luxonis/depthai-experiments/tree/master/random-scripts#oak-bandwidth-test - check your network bandwidth to make sure you don't have a bottleneck somewhere that limits the throughput.
Check the frame sizes and rate at which you are trying to send them via Ethernet. Perhaps it exceeds the available bandwidth resulting in laggy/slow execution.

Thoughts?
Jaka

Fakhrullo

jakaskerl
Hi there 🙂 as you told I checked all of them:
OAK Bandwidth - Downlink 41.3 mbps | Uplink 29.6 mbps
OAK Latency - Average latency 13.76 ms, Std: 20.9
POE Test - speed: 1000 (OK), full duplex: 1 (OK), boot mode: 3 (OK)
Blob Test - Inputs - Name: data, Type: DataType.U8F, Shape: [300, 300, 3, 1] (StorageOrder.NCHW) Outputs - Name: detection_out, Type: DataType.FP16, Shape: [7, 100, 1, 1] (StorageOrder.NCHW)
and embedding_nn.setBlobPath(blobconverter.from_zoo(name="mobilenetv2_imagenet_embedder_224x224", zoo_type="depthai", shaves=6))
is there anything incorrect?

jakaskerl

Hi Fakhrullo
The downlink and uplink are very slow (for reference: https://docs.luxonis.com/projects/api/en/latest/tutorials/low-latency/#low-latency)

Also I'm missing the part where ImgManip output is linked. Your current bandwidth requirements are 590mbit/s which is fine for gbps connection. Are you sending through anything else?

Thanks,
Jaka

Fakhrullo

jakaskerl

Hey Jaka, how r u doing? Sorry for the late response for some personal reasons I couldn't reply to you.

It's connected to RassberyPi via wifi ethernet cable, I run the test on that machine and results are good,

Bandwidth: Downlink = 647mbps, Uplink = 225mbps

Latency: Average = 2.16ms, Std: 0.5

I think there's no issue with latency or bandwidth, the problem could be there is somewhere bottleneck, what do you think bro?

jakaskerl

Hi Fakhrullo
Object counting is partially run on the host (rpi in your case) so it could be that the rpi not capable enough for your modified code and which could introduce delays.

Any chance you can add a MRE code here so we can potentially discover what the problem is?

Thanks,
Jaka