Jetson Nano best person detector model

Robotics & Edge Computing Jetson Systems Jetson Nano
,
1

Hi,
I see that there are many example on object detection for Nvidia platform, so I don’t know where to start. I would to use my Jetson Nano to detect and track only person from a webcam. Could you suggest me a models? I see that with ssd-mobilenet-v2 I can get good fps. There is a way to select just person using detectnet in Jetson-inference or I need to re-train ssd-mobilenet-v1 only for person?

Thanks

2

If you want to find only people, you may train. Otherwise, you should filter labels in your script (close other detected objects’ labels).

3

Hi,

There is a post below which emphasizes speed and accuracy between different models.

https://jonathan-hui.medium.com/object-detection-speed-and-accuracy-comparison-faster-r-cnn-r-fcn-ssd-and-yolo-5425656ae359

Yolo seems to be fast and more accurant when you compare the models. You can train you custom Yolo model. You can check this link for training custom model.

4

You can check out the production-quality PeopleNet model here: https://ngc.nvidia.com/catalog/models/nvidia:tlt_peoplenet

It was trained/pruned using Transfer Learning Toolkit and you can run it with DeepStream for the best performance.

5

Thank you so much for fast reply :) , another question, I’m trying to crop a cudaImage from jetson-inference example (detectnet.py) to get a 360 degree image from insta360 oneX2.
This is what I get from /dev/video1:

I need to crop the image half height and then concatenate the image horizontally. Using cv2 and numpy I can do it simply but I don’t know how to do using cudaCrop. I can successfully crop a cudaImage but I don’t
know how to concatenate the 2 images horizontally. I try to convert it to NumPy and then back to cudaImage and give it to detections but I got many errors .Is there a simple way to modify the input source for detectnet.py?

6

Hi @codeforge, you can use the cudaOverlay() function to compost two images into an output image. See here for example - https://github.com/dusty-nv/jetson-inference/blob/master/docs/aux-image.md#overlay

7

I will try that, thank u.

8

The performance graph shows the Nano running inference with that model at zero frames per second, so maybe not a great choice for the Nano.

9

Thanks for pointing that out - you can find a table of the data here:

https://docs.nvidia.com/metropolis/TLT/tlt-user-guide/text/purpose-built_models.html#deployment

PeopleNet is 14FPS for PeopleNet-ResNet18 and 11FPS for PeopleNet-ResNet34 on Nano.

10

That sounds much better! Thanks Dusty.