Intel Movidius NCS RPI

Image classifier and Object detection using Intel Movidius Neural Compute Stick with Raspberry Pi and ( Pi Camera or USB Web Camera)

Required Materials

1- Intel Movidius™ Neural Compute Stick

2- Development computer running Ubuntu 16.04 LTS

3- Raspberry Pi " Raspberry Pi 3 Model B Rev 1.2 has been used in this work "

4- USB Camera

5- Pi Camera

Installation and running your network with USB Camera on RPI

Please follow the installation guide provided by Intel Movidius NCS :

https://ncs-forum-uploads.s3.amazonaws.com/ncsdk/MvNC_SDK_01_07_07/NCS_Getting_Started_1.07.07.pdf

https://www.youtube.com/watch?v=f39NFuZAj6s

Pi Camera Example

1- From previous steps, you should have the following files on your Raspberry Pi :

- ncapi folder that include network's floders with each network graph file that has been compiled on your development computer.
- installed GStreamer on your Raspberry Pi

2- Install GStreamer element for the Raspberry Pi camera module (gst-rpicamsrc):

https://github.com/thaytan/gst-rpicamsrc

gst-rpicamsrc testing, run command below on your terminal, you should get a live stream preview from your PiCamera:

gst-launch-1.0 rpicamsrc bitrate=1000000 fullscreen=0 ! video/x-h264,width=640,height=480,framerate=25/1 ! filesink location=test.h264

3- Download and copy the modified python script 'stream_infearnew.py' to '../ncapi/py_examples/stream_infer/'

4- Run stream_infearnew.py

SqueezeNet Inference :

FPS ~= 9 :

Object Detection [Tiny YOLO V1] :

Compile .prototxt and corresponding .caffemodel to get the NCS graph file:

download pretrained model (.caffemodel):

 https://drive.google.com/open?id=1HBbhOwafuanewpcQ2DMpcURDveb2JNYR

By using the graph file of tiny YOLO, you will be able to get a result of object detection with [ ~5 FPS using USB 3 port and ~3 FPS using USB 2 port for Intel Movidius NCS]

Single image inference :

python3 yolo_example.py 1 ../images/person.jpg

Camera stream inference :

python3 object_detection_app.py

Acknowledgement

The author would like to thank the developers of Intel Movidius NCS, YOLONCS and gst-rpicamsrc. As well as i would like to acknowledge and express my utmost gratitude to my supervisor Professor Madya Dr. Lim Kim Chuan from Faculty of Electronics and Computer Engineering, Universiti Teknikal Malaysia Melaka (UTeM)

The equipment used in this work is provided by Machine Learning and Signal Processing Research Lab, Faculty of Electronic and Computer Engineering, Universiti Teknikal Malaysia Melaka (UTeM).