YOLOv7-tiny-vehicle-detection-on-Jetson-Nano

YOLOv7-tiny TensorRT on Jetson Nano for vehicle detection

Train model

• Download dataset from here
• Create a folder YOLOv7-tiny on your google drive
• Train model with yolov7_tiny_vehicle_detection.ipynb file.

Setup environment on Jetson Nanno B01

Build and run l4t-ml docker on Jetson Nano following below steps: I recommend to use docker instead of install packages directly on Jetson. Because of the following reasons:

• Ubuntu on Jetson Nano is not an official version released by Ubuntu, it is customize for Jetson Nano only by NVIDIA. When you try to install python package or Ubuntu package on Jetson, you will face with many compatible errors during installing process (some packages cannot find compatible version to download).
• Version update: Package version is a flustrated problem for python and Ubuntu software. When you install a package/software, it automaticall installs dependency packages. It may cause conflict with other packages dependency.
• Docker for Jetson Nano is customised by NVIDIA official. It's highly compatible with Jetson Nano and easy to set up. It reduces your headache during setup process

Log on Jetson Nano over SSH or execute Jetson Terminal Download and install docker from NVIDIA official github: https://github.com/dusty-nv/jetson-containers.git

Set docker default runtime

You need to set Docker's default-runtime to nvidia, so that the NVCC compiler and GPU are available during docker build operations. Add "default-runtime": "nvidia" to your /etc/docker/daemon.json configuration file before attempting to build the containers:

{
    "runtimes": {
        "nvidia": {
            "path": "nvidia-container-runtime",
            "runtimeArgs": []
        }
    },

    "default-runtime": "nvidia"
}

Then restart the Docker service, or reboot your system before proceeding:

sudo systemctl restart docker

You can then confirm the changes by looking under docker info

sudo docker info | grep 'Docker Root Dir'
Docker Root Dir: /mnt/docker
...
Default Runtime: nvidia

Add Swap

If you're building containers or working with large models, it's advisable to mount SWAP (typically correlated with the amount of memory in the board). Do following steps:

cd ${HOME}
git clone https://github.com/JetsonHacksNano/installSwapfile.git
cd installSwapfile
./installSwapfile.sh

After that, check swap by

free -m

Clone Docker Container Repo

sudo apt-get update && sudo apt-get install git python3-pip
git clone --depth=1 https://github.com/dusty-nv/jetson-containers
cd jetson-containers
pip3 install -r requirements.txt

Run container

I use lt4-ml container for this project since it has already contain: pytorch, tensorflow, pycuda and everything else I need to develop a deep learning application

mkdir ${HOME}/project
./run.sh -v ${HOME}/project:${HOME}/project dustynv/l4t-ml:r32.7.1

Choose to use version r32.7.1 since I'm using Jetpack 4.6

Install additional package inside docker environment

tqdm
imutils
psutil

Instal Seaborn

apt install python3-seaborn

Run Custom YOLOv7-tiny model on Jetson

Inside Docker environment

Download YOLOv7-tiny-custom model

Download pre-trained YOLOv7-tiny-custom.pt

Generate wts file from pt file

run below command to convert .pt file into .wts file

python3 gen_wts.py -w yolov7-tiny-custom.pt -o yolov7-tiny-custom.wts

Make

cd yolov7/
mkdir build
cd build
cp ../../yolov7-tiny-custom.wts .
cmake ..
make 

Build TensorRT engine file

./yolov7 -s yolov7-tiny-custom.wts  yolov7-tiny-custom.engine t

Python Object Detection Use app.py to do inference on video file

python3 app.py

If you have custom model, make sure to update categories as per your classes in yolovDet.py