How to install Yolo v4 on Jetson Xavier NX

In this tutorial, we will see how to use Jetson Xavier NX with YOLO v4 and darknet. In the past, we have already talked about the Nvidia system module and for comparison or simple curiosity, I invite you to read and watch the YOLO V3 – Install and run Yolo on Nvidia Jetson Nano (with GPU) video.

Jetson Xavier NX technical details

When dealing with Computer vision, the first most important feature is the graphics card:

GPU: NVIDIA Volta with 384 CUDA Cores, 48 Tensor Cores

Given its size, the power consumption, and the cost that on the Nvidia store is $ 399 make it an excellent product for specific projects, especially when we cannot use a Supercomputer. I show you a table with the most important technical details but I recommend you visit the official Jetson Xavier NX Nvidia page

[bg_collapse_level2 view=”button-blue” color=”#fffff” icon=eye” expand_text=”Show More technical details” collapse_text=”Show less”]

GPU	384-core NVIDIA Volta™ GPU with 48 Tensor Cores
CPU	6-core NVIDIA Carmel ARM®v8.2 64-bit CPU 6 MB L2 + 4 MB L3
Memory	8 GB 128-bit LPDDR4x 51.2GB/s
Storage	microSD (Card not included)
VIDEO ENCODER	2x 4Kp30 | 6x 1080p 60 | 14x 1080p30 (H.265/H.264)
VIDEO DECODER	2x 4Kp60 | 4x 4Kp30 | 12x 1080p60 | 32x 1080p30 (H.265) 2x 4Kp30 | 6x 1080p60 |16x 1080p30 (H.264)
Connectivity	Gigabit Ethernet, M.2 Key E (WiFi/BT included), M.2 Key M (NVMe)
Camera	2x MIPI CSI-2 D-PHY lanes
Display	HDMI and DP
Others	GPIOs, I2C, I2S, SPI, UART
Mechanical	69.6 mm x 45 mm | 260-pin SO-DIMM connector
AI Perf	21 TOPS
POWER	10W | 15W
PCIe	1 x1 (PCIe Gen3) + 1 x4 (PCIe Gen4)
CSI Camera	Up to 6 cameras (24 via virtual channels) | 14 lanes (3×4 or 6×2) MIPI CSI-2 | D-PHY 1.2 (up to 30 Gbps)
DL Accelerator	2x NVDLA Engines
VISION ACCELERATOR	7-Way VLIW Vision Processor
Networking	10/100/1000 BASE-T Ethernet

[/bg_collapse_level2]

Install Yolo v4 on Jetson Xavier NX

The goal of this tutorial is to install YOLOv4 on Nvidia Jetson Xavier NX and run it in real-time. I will not delve into the aspects of installing the operating system in the MicroSD but I will leave you some links to find the correct JetPack.

Where can i find the JetPack and how do i use it?

This is the JetPack Archive page with the complete archive, I recommend you download the latest available.
Just downloaded you have to write the image in the MicroSD card and this Getting Started With Jetson Xavier NX Developer Kit page can help you.

If the MicroSD card is ready you can put it in the device and follow these instructions How to Install JetPack for the first installation, this is the result at startup.

Download, Install and Run Darknet

For this object detection project, we will use Darknet which is the framework on which to use the Yolo v4 deep learning algorithm. Now open the terminal and run the commands step by step.

1 Update
Update everything

sudo apt-get update

2 EXPORT CUDA PATH
After checking for updates we have to copy, one at a time, the commands for export CUDA path.

export PATH=/usr/local/cuda-10.2/bin${PATH:+:${PATH}}

export LD_LIBRARY_PATH=/usr/local/cuda-10.2/lib64${LD_LIBRARY_PATH:+:${LD_LIBRARY_PATH}}

As you can see from the image, the command produces no output

3 INSTALL DARKNET
We download darknet from GitHub always through the terminal

git clone https://github.com/AlexeyAB/darknet

When the download finishes you will find a new folder called Darknet.

Access to the folder through the terminal

cd darknet

If it looks like it is in the terminal, everything went as it should and you are inside the folder.

Now let’s run the commands to download the model.

wget https://github.com/AlexeyAB/darknet/releases/download/darknet_yolo_v4_pre/yolov4-csp.weights

wget https://raw.githubusercontent.com/AlexeyAB/darknet/master/cfg/yolov4-csp.cfg

4 ENABLE GPU
This is the most important step, we need to enable the GPU and then take advantage of the full power of the graphics card of this device. Also from the terminal execute the command

sudo vi Makefile

through the VIM editor you need GPU, CUDNN, OpenCV. The result must be as in the image.

Now I have pressed Esc + shift + z + z to save.

5 COMPILE
The last step, we need to fill out. Always in the darknet folder from the terminal, we execute the command

make

Hopefully, it should fill out in a few minutes and we’re ready to test.

Run YOLO v4 from a video

To test with YOLO v4 I downloaded a video called jump_water.mp4 and put it in the Darknet folder. Now run this command to start processing

./darknet detector demo cfg/coco.data cfg/yolov4-csp.cfg yolov4-csp.weights -ext_output jump_water.mp4

After loading yolo model on the memory, this is the result

Run YOLO v4 from a camera

Now let’s try to execute the command to start Yolo in realtime with the camera

./darknet detector demo cfg/coco.data cfg/yolov4-csp.cfg yolov4-csp.weights -c 0