Face Recognition Project using Jetson Nano Board.

[singh.bharat.bsc]

Hello,
I am a student currently pursuing B. Tech. CSE. I am making a project of Face Recognition which, of course, recognize faces in a given room, and then categorize each face in different categories. For example: Manager, Client, Worker, Visitor, etc. Also if any intruder comes in the room, then it will raise an alarm and alert the security team. I have designed the software and tested it on my laptop, but it was not that fast and powerful to do such complex things. And that’s why i am here to find a suitable board for my project.

I have found the NVIDIA Jetson Nano board perfect for this kind of work. ( I think so because the specs says so…) but i don’t know that whether it can handle this complex job in which it have to analyze each frame of a live feed so fast, that it does not lag. So, my questions are:

Whether it can process these kind of things…?

How can i connect it to the internet without connecting the board with the Ethernet ?(through WiFi… because i have seen that this board does not have WiFi in-built)

Do you think that this board is enough…? or do you recommend me another one…?

Thanks.

[dusty_nv]

Hi singh.bharat, you might be interested in this Smart Doorbell project for Nano which performs face recognition: https://medium.com/@ageitgey/build-a-hardware-based-face-recognition-system-for-150-with-the-nvidia-jetson-nano-and-python-a25cb8c891fd

You could achieve better performance using TensorRT for the DNN inferencing, but that is a good start.
For WiFi, you could check USB dongle or M.2 Key-E modules from the Supported Components List.

[KoDaniel]

I recently did my own face recognition task on a jetson nano for my Masters degree. First I started with the Doorbell too but it was not satisfying. The Dlib models worked too slow on the Jetson in my opinion.

For face detection I adapted this TRT-MTCNN https://github.com/PKUZHOU/MTCNN_FaceDetection_TensorRT to my Jetson.
For Recognition I use:

1. Dlib face embeddings as feature extraction (dlib_face_recognition_model)
2. dlib SVMs for classification

I can run the Pipeline on camera images or on images from disk. I trained and tested it on my family and I created a BBT dataset for training/testing. Its trained on 7 Classes and it works quite well. I can detect and classify an image with 10 faces with ~10 FPS (see attachement). In my implementation adding more Classes will decrease the speed. I wouldn’t use more than ~10 classes.
here is my Project https://github.com/Daniel595/Jetson_nano_face_recognition
and testdata https://github.com/Daniel595/testdata