Nano-Yolov7-ESP32

How to setup and perform Yolov7 inference object detection on Jetson Nano, using ESP32 wifi camera.

---

Notes

• Current versions of both Yolov5, and Yolov7 state that Python 3.7 or higher are required
• Jetson Nano struggles in using any Python version that is different than the installed ones (Python 3.6.9)
• In this project we are showing the lengthy setup required to run Yolov7 on the Nano, without changing from Python 3.6.9
• For this project you need a MicroSD card, Jetson Nano, and an ESP-32 Camera Board

---

Installation

---

First: Download and flash Jetpack 4.6 OS on 64+gb micro SD card

• Go to this page and download the correct version of Jetpack for your nano
• Unzip the file
• Now download Balena Etcher which will allow you to flash the unzipped file onto your MicroSD card.
• Plug the MicroSD card into your Nano's MicroSD slot, and run the OS setup

Now: Create addition virtual swap memory to help with performance by running the following shell commands

# Disable ZRAM:
sudo systemctl disable nvzramconfig

# Create 4GB swap file
sudo fallocate -l 4G /mnt/4GB.swap
sudo chmod 600 /mnt/4GB.swap
sudo mkswap /mnt/4GB.swap

# Append the following line to /etc/fstab
sudo su
echo "/mnt/4GB.swap swap swap defaults 0 0" >> /etc/fstab
exit

# REBOOT!

---

After rebooting run the following shell commands:

sudo apt-get update
sudo apt-get upgrade

Now we need to install the pip Python package manager by running the following in the shell

wget https://bootstrap.pypa.io/get-pip.py

python3 get-pip.py

---

Pull Pytorch for Jetson Nano from Nvidia, and install some dependencies before installing PyTorch

wget https://nvidia.box.com/shared/static/p57jwntv436lfrd78inwl7iml6p13fzh.whl -O torch-1.8.0-cp36-cp36m-linux_aarch64.whl

sudo apt-get install libopenblas-base libopenmpi-dev libomp-dev

Now, in the same directory, pip install Cython and PyTorch

    python3 -m pip install Cython
    python3 -m pip install torch-1.7.0-cp36-cp36m-linux_aarch64.whl

Note: Installing PyTorch can take between 15-30 minutes

Now check to see if PyTorch is installed, with GPU enabled

python3
>>> import torch
>>>torch.cuda.is_available()

If we get True, we can move on to the TorchVision Installation

Before, we need to install some more linux dependencies

sudo apt-get install libjpeg-dev zlib1g-dev libpython3-dev libavcodec-dev libavformat-dev libswscale-dev

Now we need to clone the torchvision repo, using v0.9.0 because it is compatible with PyTorch v1.8, which we just installed

git clone --branch v0.9.0 https://github.com/pytorch/vision torchvision

cd torchvision

python3 setup.py install

Note: Torchvision can take up to 40 minutes to install

---

Now, we install some linux and python packages, and set the an environment variable:

sudo apt install libfreetype6-dev
export OPENBLAS_CORETYPE=ARMV8 python3
sudo apt-get install python3-seaborn

sudo apt-get install python3-pil
python3 -m pip install future
pip3 install 'pillow<7'
pip3 install tdqm

Now that we have setup and installed all dependencies required to run Yolov7, clone the repo

git clone https://github.com/WongKinYiu/yolov7

cd yolov7

---

Before running Yolov7's inference, we must setup the ESP32 camera:

• On a different machine, install Arduino IDE
• Now clone this repository, that has the code we need to flash the ESP32

git clone https://github.com/arkhipenko/esp32-cam-mjpeg.git
cd esp32-cam-mjpeg

Before flashing, open and setup ArduinoIDE so that it is compatible with our board

• Under the file dropdown menu, to preferences, and paste this link in the field that says 'Additional Board Manager URLs' https://raw.githubusercontent.com/espressif/arduino-esp32/gh-pages/package_esp32_index.json
• Now click on tools, Board, and then Board Manager
• Search for esp32 in the search bar, and install the first package that appears
• Plug in your ESP32 cam via microUSB to your computer
• Click on Tools, the click on Board, and select 'AI Thinker ESP32-CAM'
• Click on Tools, then choose any available Ports
• Open a terminal and change directories to the esp32-cam-mjpeg repo we just downloaded
• Open esp32_camera_mjpeg.ino file with Arduino IDE
• In the file, comment out the line that says

#define CAMERA_MODEL_ESP_EYE

• And uncomment the line that says

//#define CAMERA_MODEL_AI_THINKER

• In the same directory make a file called "home_wifi_multi.h", drop in the following code, and save. Replace the strings with your Wifi station name, and password

#define SSID1 "replace with your wifi ssid"
#define PWD1 "replace your wifi password"

• Now on ArduinoIDE upload the code, and open the Serial Monitor, setting it to 115200 baud
• Press the reset button on the ESP32 cam, and you should see an output similar to the following:

192.168.1.42
Stream Link: http://192.168.1.42/mjpeg/1

• This is the stream link we are going to use on the Nano, with Yolov7

---

On the Jetson Nano, change directories to the yolov7 repo we cloned earlier

• Run the following command

python3 detect.py --source http://192.168.1.42/mjpeg/1

Obviously, replace this stream link with the one you obtained from ArduinoIDE

Now, we have run Yolov7 inference object detection on the Jetson Nano, using the ESP32 wifi camera.

• The following is a video capture of me running Yolov7 on my Jetson Nano https://www.youtube.com/watch?v=IZY73M8Pzbk