21、Jetson Xavier Nx使用Centerface+Arcface进行人脸识别、认证、比对。并且使用USB转TTL进行串口通信

基本思想：使用Jetson Xavier NX开发板进行人脸识别和认证比对，首先进行Jetson Xavier Nx的环境搭建，然后在进行人脸检测模型选择，最后进行移植Jetson Xavier NX开发板

一、配置Jetson Xavier NX开发板环境，因为系统镜像已将安装，这里就不在详细叙述了；

GitHub - linghu8812/tensorrt_inference 版本用https://github.com/linghu8812/tensorrt_inference/tree/demo/0.1.0

安装的mxnet==1.5.0 onnx==1.5.0 python3.6 建议anconda环境转模型

因此在系统镜像的基础上，进行必要的软件安装百度云盘提供了两whl文件

链接：https://pan.baidu.com/s/11-E50MR-7Nup-D8kBdynnA
提取码：53hh

(1)、安装Pytorch-1.10.0

# install the dependencies (if not already onboard)
nvidia@nvidia-desktop:~$ sudo apt-get install python3-pip libjpeg-dev libopenblas-dev libopenmpi-dev libomp-dev
nvidia@nvidia-desktop:~$ sudo -H pip3 install future
nvidia@nvidia-desktop:~$ sudo pip3 install -U --user wheel mock pillow
nvidia@nvidia-desktop:~$ sudo -H pip3 install testresources
# upgrade setuptools 47.1.1 -> 58.3.0
nvidia@nvidia-desktop:~$ sudo -H pip3 install --upgrade setuptools
nvidia@nvidia-desktop:~$ sudo -H pip3 install Cython
# install gdown to download from Google drive
nvidia@nvidia-desktop:~$ sudo -H pip3 install gdown
# download the wheel
nvidia@nvidia-desktop:~$ gdown https://drive.google.com/uc?id=1TqC6_2cwqiYacjoLhLgrZoap6-sVL2sd
# install PyTorch 1.10.0
nvidia@nvidia-desktop:~$ sudo -H pip3 install torch-1.10.0a0+git36449ea-cp36-cp36m-linux_aarch64.whl
# clean up
nvidia@nvidia-desktop:~$ rm torch-1.10.0a0+git36449ea-cp36-cp36m-linux_aarch64.whl

(2)、安装Torchvison 0.11.0

Used with PyTorch 1.10.0
# the dependencies
nvidia@nvidia-desktop:~$ sudo apt-get install libjpeg-dev zlib1g-dev libpython3-dev
nvidia@nvidia-desktop:~$ sudo apt-get install libavcodec-dev libavformat-dev libswscale-dev
nvidia@nvidia-desktop:~$ sudo pip3 install -U pillow
# install gdown to download from Google drive, if not done yet
nvidia@nvidia-desktop:~$ sudo -H pip3 install gdown
# download TorchVision 0.11.0
nvidia@nvidia-desktop:~$ gdown https://drive.google.com/uc?id=1C7y6VSIBkmL2RQnVy8xF9cAnrrpJiJ-K
# install TorchVision 0.11.0
nvidia@nvidia-desktop:~$ sudo -H pip3 install torchvision-0.11.0a0+fa347eb-cp36-cp36m-linux_aarch64.whl
nvidia@nvidia-desktop:~$ pip3 install pycuda
# clean up
nvidia@nvidia-desktop:~$ rm torchvision-0.11.0a0+fa347eb-cp36-cp36m-linux_aarch64.whl

jtop查看cpu和gpu、内存利用率

nvidia@nvidia-desktop:~$ sudo apt-get install git cmake
nvidia@nvidia-desktop:~$ sudo apt-get install python3-dev
nvidia@nvidia-desktop:~$ sudo apt-get install libhdf5-serial-dev hdf5-tools
nvidia@nvidia-desktop:~$ sudo apt-get install libatlas-base-dev gfortran
nvidia@nvidia-desktop:~$ sudo -H pip3 install -U jetson-stats
nvidia@nvidia-desktop:~$ sudo jtop

显示的Jetson Xavier NX开发板的使用率

测试一下安装库

nvidia@nvidia-desktop:~$ python3
Python 3.6.9 (default, Jan 26 2021, 15:33:00)
[GCC 8.4.0] on linux
Type "help", "copyright", "credits" or "license" for more information.
>>> import torch
>>> torch.__version__
'1.10.0a0+git36449ea'
>>> import torchvision
>>> torchvision.__version__
'0.11.0a0+fa347eb'
>>> import tensorrt
>>> tensorrt.__version__
'7.1.3.0'
>>> import pycuda
>>>

整个Python3库的版本信息

nvidia@nvidia-desktop:~$ pip3 list
DEPRECATION: The default format will switch to columns in the future. You can use --format=(legacy|columns) (or define a format=(legacy|columns) in your pip.conf under the [list] section) to disable this warning.
appdirs (1.4.4)
apt-clone (0.2.1)
apturl (0.5.2)
asn1crypto (0.24.0)
beautifulsoup4 (4.6.0)
blinker (1.4)
Brlapi (0.6.6)
certifi (2018.1.18)
chardet (3.0.4)
cryptography (2.1.4)
cupshelpers (1.0)
cycler (0.11.0)
Cython (0.29.24)
dataclasses (0.8)
decorator (4.1.2)
defer (1.0.6)
distro-info (0.18ubuntu0.18.04.1)
feedparser (5.2.1)
filelock (3.3.2)
future (0.18.2)
gdown (4.2.0)
graphsurgeon (0.4.5)
html5lib (0.999999999)
httplib2 (0.9.2)
idna (2.6)
jetson-stats (3.1.1)
Jetson.GPIO (2.0.8)
keyring (10.6.0)
keyrings.alt (3.0)
language-selector (0.1)
launchpadlib (1.10.6)
lazr.restfulclient (0.13.5)
lazr.uri (1.0.3)
louis (3.5.0)
lxml (4.2.1)
macaroonbakery (1.1.3)
Mako (1.1.5)
MarkupSafe (2.0.1)
matplotlib (2.1.1)
numpy (1.19.5)
oauth (1.0.1)
oauthlib (2.0.6)
olefile (0.45.1)
onboard (1.4.1)
PAM (0.4.2)
pandas (0.22.0)
patsy (0.4.1+dev)
pbr (5.7.0)
Pillow (8.4.0)
pip (9.0.1)
protobuf (3.0.0)
pycairo (1.16.2)
pycrypto (2.6.1)
pycuda (2021.1)
pycups (1.9.73)
pygobject (3.26.1)
PyICU (1.9.8)
PyJWT (1.5.3)
pymacaroons (0.13.0)
PyNaCl (1.1.2)
pyparsing (2.2.0)
pyRFC3339 (1.0)
PySocks (1.7.1)
python-apt (1.6.5+ubuntu0.2)
python-dateutil (2.6.1)
python-debian (0.1.32)
pytools (2021.2.9)
pytz (2018.3)
pyxdg (0.25)
PyYAML (3.12)
requests (2.18.4)
requests-unixsocket (0.1.5)
scipy (0.19.1)
seaborn (0.8.0)
SecretStorage (2.3.1)
setuptools (58.5.3)
simplejson (3.13.2)
six (1.11.0)
ssh-import-id (5.7)
system-service (0.3)
systemd-python (234)
tensorrt (7.1.3.0)
testresources (2.0.1)
torch (1.10.0a0+git36449ea)
torchvision (0.11.0a0+fa347eb)
tqdm (4.62.3)
typing-extensions (3.10.0.2)
ubuntu-drivers-common (0.0.0)
uff (0.6.9)
unity-scope-calculator (0.1)
unity-scope-chromiumbookmarks (0.1)
unity-scope-colourlovers (0.1)
unity-scope-devhelp (0.1)
unity-scope-firefoxbookmarks (0.1)
unity-scope-manpages (0.1)
unity-scope-openclipart (0.1)
unity-scope-texdoc (0.1)
unity-scope-tomboy (0.1)
unity-scope-virtualbox (0.1)
unity-scope-yelp (0.1)
unity-scope-zotero (0.1)
urllib3 (1.22)
urwid (2.0.1)
wadllib (1.3.2)
webencodings (0.5)
wheel (0.30.0)
xkit (0.0.0)
zope.interface (4.3.2)

(3)配置一下Jetson Xavier NX开发板的vnc连接

sudo vim /usr/share/glib-2.0/schemas/org.gnome.Vino.gschema.xml

添加内容（不要和其它key重叠）添加在

  </schema>
</schemalist>

之前；具体内容

<key name='enabled' type='b'><summary>Enable remote access to the desktop</summary><description>If true, allows remote access to the desktop via the RFBprotocol. Users on remote machines may then connect to thedesktop using a VNC viewer.</description><default>true</default>
</key>

执行

nvidia@nvidia-desktop:~$ sudo glib-compile-schemas /usr/share/glib-2.0/schemas
nvidia@nvidia-desktop:~$

搜索栏设置搜索一下

然后图形化界面设置一下

打开之后，点击ADD，然后新建一个程序。 Name 设置为Vino, Command设置为/usr/lib/vino/vino-server . Comment设置为VNC Server。设置好之后保存即可。

终端输入以下指令来关闭VNC连接加密使用时候设置全部设置true

gsettings set org.gnome.Vino require-encryption false
gsettings set org.gnome.Vino prompt-enabled false

测试结果 vnc viewer连接Jetson Xavier NX开发板

以后每次使用vnc的前提下，需要先进入图形化界面之后，才能使用VNC连接

环境基本搭建完毕~~~~~

二、先在PC的笔记本上测试一下需要使用的人脸检测算法，然后在移植jetson nano nx ；

首先测试一下mtcnn的识别速度和精度，然后再试一下Centerface的识别速度和精度~

（1）下载源码首先先对mtcnn测试，这个之前其实用过 15、JetBot进行目标跟踪及人脸匹配跟踪_sxj731533730-CSDN博客

git clone https://github.com/jkjung-avt/tensorrt_demos.git

在环境配置好的基础上，50、ubuntu18.04/20.04进行TensorRT环境搭建和YOLO5部署(含安装vulkan）_sxj731533730-CSDN博客

（2）将库文件

/home/ubuntu/NVIDIA_CUDA-11.1_Samples/TensorRT-7.2.2.3/lib

文件拷贝到

/home/ubuntu/tensorrt_demos/mtcnn

（3）将头文件

/home/ubuntu/NVIDIA_CUDA-11.1_Samples/TensorRT-7.2.2.3/include

拷贝到

/home/ubuntu/tensorrt_demos/common

五、编译mtcnn文件生成mtcnn_tengine,然后生成tensorrt权重文件

ubuntu@ubuntu:~/tensorrt_demos/mtcnn$ make
../common/NvInfer.h:3250:22: note: declared here3250 | class TRT_DEPRECATED IRNNv2Layer : public ILayer|                      ^~~~~~~~~~~
../common/NvInfer.h:5662:85: warning: ‘IPluginLayer’ is deprecated [-Wdeprecated-declarations]5662 |  inputs, int32_t nbInputs, IPluginExt& plugin) TRTNOEXCEPT = 0;|                                                              ^../common/NvInfer.h:3454:22: note: declared here3454 | class TRT_DEPRECATED IPluginLayer : public ILayer|                      ^~~~~~~~~~~~
In file included from create_engines.cpp:30:
../common/NvCaffeParser.h:108:62: warning: ‘DimsNCHW’ is deprecated [-Wdeprecated-declarations]108 |     virtual nvinfer1::DimsNCHW getDimensions() TRTNOEXCEPT = 0;|                                                              ^
In file included from create_engines.cpp:29:
../common/NvInfer.h:346:22: note: declared here346 | class TRT_DEPRECATED DimsNCHW : public Dims4|                      ^~~~~~~~
Linking: create_engines
ubuntu@ubuntu:~/tensorrt_demos/mtcnn$ ./create_engines
Building det1.engine (PNet), maxBatchSize = 1
Building TensorRT engine in FP32 mode...
WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
Building det2.engine (RNet), maxBatchSize = 256
Building TensorRT engine in FP32 mode...
WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
Building det3.engine (ONet), maxBatchSize = 64
Building TensorRT engine in FP32 mode...
WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1Verifying engines...
WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
Bindings for det1 after deserializing:Input  0: data, 3x710x384Output 1: conv4-2, 4x350x187Output 2: prob1, 2x350x187
WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
Bindings for det2 after deserializing:Input  0: data, 3x24x24Output 1: conv5-2, 4x1x1Output 2: prob1, 2x1x1
WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
Bindings for det3 after deserializing:Input  0: data, 3x48x48Output 1: conv6-2, 4x1x1Output 2: conv6-3, 10x1x1Output 3: prob1, 2x1x1
Done.

然后生成/home/ubuntu/tensorrt_demos/setup.py

det1.engine
det2.engine
det3.engine

六、修改 /home/ubuntu/tensorrt_demos/setup.py

'/home/ubuntu/NVIDIA_CUDA-11.1_Samples/TensorRT-7.2.2.3/lib',  # for my x86_64 PC
'/home/ubuntu/NVIDIA_CUDA-11.1_Samples/TensorRT-7.2.2.3/include',  # for my x86_64 PC

然后make一下

buntu@ubuntu:~/tensorrt_demos$ make
x86_64-linux-gnu-g++ -pthread -shared -Wl,-O1 -Wl,-Bsymbolic-functions -Wl,-Bsymbolic-functions -Wl,-z,relro -g -fwrapv -O2 -Wl,-Bsymbolic-functions -Wl,-z,relro -g -fwrapv -O2 -g -fstack-protector-strong -Wformat -Werror=format-security -Wdate-time -D_FORTIFY_SOURCE=2 build/temp.linux-x86_64-3.8/pytrt.o -L/usr/local/cuda/lib64 -L/home/ubuntu/NVIDIA_CUDA-11.1_Samples/TensorRT-7.2.2.3/lib -L/usr/local/lib -lnvinfer -lcudnn -lcublas -lcudart_static -lnvToolsExt -lcudart -lrt -o /home/ubuntu/tensorrt_demos/pytrt.cpython-38-x86_64-linux-gnu.so
rm -rf build

测试一下mtcnn 【需要修改一下识别尺寸 480 640 与centerface做对比】

ubuntu@ubuntu:~/tensorrt_demos$ python3 trt_mtcnn.py --image ../CenterFace/prj-python/000388.jpg

测试速度 20fps 测试效果图

七、下载Centerface源码，同时安装pycuda

git clone https://github.com/Star-Clouds/CenterFace
pip3 install pycuda

然后进入超级权限状态下针对centerface tensorrt

sudo su
export PATH=/usr/local/cuda-11.1/bin:/usr/local/cuda/bin:$PATH 　　
pip install pycuda

遇到问题一应该是模型的版本和我系统的Tensorrt版本不一致无法解析模型

root@ubuntu:/home/ubuntu/CenterFace/prj-tensorrt# python3 demo.py
[TensorRT] ERROR: coreReadArchive.cpp (41) - Serialization Error in verifyHeader: 0 (Version tag does not match. Note: Current Version: 96, Serialized Engine Version: 87)
[TensorRT] ERROR: INVALID_STATE: std::exception
[TensorRT] ERROR: INVALID_CONFIG: Deserialize the cuda engine failed.
Traceback (most recent call last):File "demo.py", line 30, in <module>test_image_tensorrt()File "demo.py", line 13, in test_image_tensorrtdets, lms = centerface(frame, h, w, threshold=0.35)File "/home/ubuntu/CenterFace/prj-tensorrt/centerface.py", line 21, in __call__return self.inference_tensorrt(img, threshold)File "/home/ubuntu/CenterFace/prj-tensorrt/centerface.py", line 73, in inference_tensorrtcontext = engine.create_execution_context()
AttributeError: 'NoneType' object has no attribute 'create_execution_context'
root@ubuntu:/home/ubuntu/CenterFace/prj-tensorrt# python3 demo.py
[TensorRT] ERROR: coreReadArchive.cpp (41) - Serialization Error in verifyHeader: 0 (Version tag does not match. Note: Current Version: 96, Serialized Engine Version: 87)
[TensorRT] ERROR: INVALID_STATE: std::exception
[TensorRT] ERROR: INVALID_CONFIG: Deserialize the cuda engine failed.
Traceback (most recent call last):File "demo.py", line 30, in <module>test_image_tensorrt()File "demo.py", line 13, in test_image_tensorrtdets, lms = centerface(frame, h, w, threshold=0.35)File "/home/ubuntu/CenterFace/prj-tensorrt/centerface.py", line 21, in __call__return self.inference_tensorrt(img, threshold)File "/home/ubuntu/CenterFace/prj-tensorrt/centerface.py", line 73, in inference_tensorrtcontext = engine.create_execution_context()
AttributeError: 'NoneType' object has no attribute 'create_execution_context'

重新下载一下模型转化代码

https://github.com/linghu8812/tensorrt_inference

使用Clion生成一下模型（centerface.onnx 来自官方代码）

（1）使用onnx导出模型设置为640 480大小输入

import onnx
import math
import argparseparser = argparse.ArgumentParser(description='Export CenterFace ONNX')
parser.add_argument('--pretrained', help='pretrained centerface model', default='./centerface.onnx', type=str)
parser.add_argument('--input_shape', nargs='+', default=[1, 3, 480, 640], type=int, help='input shape.')
parser.add_argument('--onnx', help='onnx model', default='./centerface_480_640.onnx', type=str)
args = parser.parse_args()model = onnx.load_model(args.pretrained)
input_shape = args.input_shaped = model.graph.input[0].type.tensor_type.shape.dim
rate = (input_shape[2] / d[2].dim_value, input_shape[3] / d[3].dim_value)
print("rate: ", rate)
d[0].dim_value = input_shape[0]
d[2].dim_value = int(d[2].dim_value * rate[0])
d[3].dim_value = int(d[3].dim_value * rate[1])
for output in model.graph.output:d = output.type.tensor_type.shape.dimd[0].dim_value = input_shape[0]d[2].dim_value = int(d[2].dim_value * rate[0])d[3].dim_value = int(d[3].dim_value * rate[1])onnx.save_model(model, args.onnx)

（2）修改代码config.yaml配置文件

CenterFace:onnx_file:     "/home/ubuntu/tensorrt_inference/CenterFace/centerface_480_640.onnx"engine_file:   "/home/ubuntu/tensorrt_inference/CenterFace/centerface.trt"BATCH_SIZE:    1INPUT_CHANNEL: 3IMAGE_WIDTH:   640IMAGE_HEIGHT:  480obj_threshold: 0.5nms_threshold: 0.45

同时修改CMakelists.txt文件中的原内容

set(TENSORRT_ROOT /usr/src/tensorrt/)

修改之后的内容为

set(TENSORRT_ROOT /home/ubuntu/NVIDIA_CUDA-11.1_Samples/TensorRT-7.2.2.3)

创建build文件夹下，执行cmake& make执行命令

ubuntu@ubuntu:~/tensorrt_inference/CenterFace/build$ ./CenterFace_trt ../config.yaml ../samples/
loading filename from:/home/ubuntu/tensorrt_inference/CenterFace/centerface.trt
[11/12/2021-10:50:02] [W] [TRT] TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
deserialize done
[11/12/2021-10:50:02] [W] [TRT] TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
binding0: 3686400
binding1: 76800
binding2: 153600
binding3: 153600
binding4: 768000
Processing: ../samples//worlds-largest-selfie.jpg
prepareImage
prepare image take: 31.9722 ms.
host2device
execute
Inference take: 1.60191 ms.
execute success
device2host
post process
Post process take: 2583.01 ms.
../samples//worlds-largest-selfie_.jpg
Processing: ../samples//test4.jpg
prepareImage
prepare image take: 6.97756 ms.
host2device
execute
Inference take: 1.42854 ms.
execute success
device2host
post process
Post process take: 313.282 ms.
../samples//test4_.jpg
Processing: ../samples//test5.jpg
prepareImage
prepare image take: 3.43697 ms.
host2device
execute
Inference take: 1.39913 ms.
execute success
device2host
post process
Post process take: 72.2356 ms.
../samples//test5_.jpg
Processing: ../samples//test3.jpg
prepareImage
prepare image take: 3.55852 ms.
host2device
execute
Inference take: 1.4428 ms.
execute success
device2host
post process
Post process take: 265.396 ms.
../samples//test3_.jpg
Processing: ../samples//test1.jpg
prepareImage
prepare image take: 3.9262 ms.
host2device
execute
Inference take: 1.40752 ms.
execute success
device2host
post process
Post process take: 475.307 ms.
../samples//test1_.jpg
Processing: ../samples//test2.jpg
prepareImage
prepare image take: 4.02272 ms.
host2device
execute
Inference take: 1.40933 ms.
execute success
device2host
post process
Post process take: 832.401 ms.
../samples//test2_.jpg
Average processing time is 767.369ms

或者使用IDE工开发工具编译和执行

生成日志

/home/ubuntu/tensorrt_inference/CenterFace/cmake-build-debug/CenterFace_trt /home/ubuntu/tensorrt_inference/CenterFace/config.yaml /home/ubuntu/tensorrt_inference/CenterFace/sample
----------------------------------------------------------------
Input filename:   /home/ubuntu/tensorrt_inference/CenterFace/centerface.onnx
ONNX IR version:  0.0.4
Opset version:    9
Producer name:    pytorch
Producer version: 1.2
Domain:
Model version:    0
Doc string:
----------------------------------------------------------------
[11/06/2021-16:53:16] start building engine
[11/06/2021-16:53:16] [W] [TRT] onnx2trt_utils.cpp:220: Your ONNX model has been generated with INT64 weights, while TensorRT does not natively support INT64. Attempting to cast down to INT32.
[W] [TRT] Half2 support requested on hardware without native FP16 support, performance will be negatively affected.
[W] [TRT] TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
[11/06/2021-16:53:17] [11/06/2021-16:53:26] [I] [TRT] Some tactics do not have sufficient workspace memory to run. Increasing workspace size may increase performance, please check verbose output.
[11/06/2021-16:53:57] [I] [TRT] Detected 1 inputs and 4 output network tensors.
[11/06/2021-16:53:57] build engine done
[W] [TRT] TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
writing engine file...
save engine file done
[11/06/2021-16:53:57] binding0: 4915200
binding1: 102400
binding2: 204800
binding3: 204800
binding4: 1024000
Average processing time is -nanms
[W] [TRT] TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1Process finished with exit code 0

生成结果c++

（3）因为生成模型是480 640 的模型，主要原因是MTCNN也是这样的大小进行检测的，为了对比，选择使用哪个模型进行关键点检测和比对

测试结果注意测试结果要在root权限运行，貌似需要调用NVCC需要系统权限

这里发现一个问题，如果循环执行该代码时，会遇到显存不够的情况，所以修改一下Centerface.py代码

识别速度还是蛮快的 RTX2060

/usr/bin/python3.8 /home/ubuntu/CenterFace/prj-tensorrt/demo.py
[TensorRT] WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
[TensorRT] WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
gpu times =  2.311 ms
(1, 1, 120, 160) (1, 2, 120, 160) (1, 2, 120, 160) (1, 10, 120, 160)
count =  1
gpu times =  2.124 ms
(1, 1, 120, 160) (1, 2, 120, 160) (1, 2, 120, 160) (1, 10, 120, 160)
count =  1
gpu times =  2.316 ms
(1, 1, 120, 160) (1, 2, 120, 160) (1, 2, 120, 160) (1, 10, 120, 160)
count =  1
gpu times =  2.833 ms
(1, 1, 120, 160) (1, 2, 120, 160) (1, 2, 120, 160) (1, 10, 120, 160)
count =  1
gpu times =  2.516 ms
(1, 1, 120, 160) (1, 2, 120, 160) (1, 2, 120, 160) (1, 10, 120, 160)
count =  1
gpu times =  4.811 ms
(1, 1, 120, 160) (1, 2, 120, 160) (1, 2, 120, 160) (1, 10, 120, 160)
count =  1
gpu times =  2.261 ms
(1, 1, 120, 160) (1, 2, 120, 160) (1, 2, 120, 160) (1, 10, 120, 160)
count =  1
gpu times =  2.168 ms
(1, 1, 120, 160) (1, 2, 120, 160) (1, 2, 120, 160) (1, 10, 120, 160)
count =  1

相比而言centerface的识别速度和精度比mtcnn更胜一筹，开始写人脸对齐和人证比对的代码；下图为运行CenterFace-TensorRT过程中的显存使用率

八、生成arface的tengine模型，安装mxnet的gpu版和下载预训练模型进行转onnx （我的驱动版本是cuda-11.2）

（1）安装mxnet库之后，/home/ubuntu/tensorrt_inference/arcface/inference.py 可以测试在模型转完之后，测试下载的预训练模型是否可用（非必要操作）

pip3 install -i https://pypi.tuna.tsinghua.edu.cn/simple mxnet-cu112

我将下载的mobilefacenet-res2-6-10-2-dim512文件夹放在了/home/ubuntu/tensorrt_inference/arcface文件中，然后使用arcface_retinaface_mxnet2onnx/mxnet2onnx_demo.py at master · zheshipinyinMc/arcface_retinaface_mxnet2onnx · GitHub 中的arcface_retinaface_mxnet2onnx.py脚本的以下三个函数对mxnet模型进行onnx生成

mxnet2onnx_test() #==mxnet2onnxonnx_modify_demo() #===onnx修改===onnx_inferred_demo() #===onnx前向推导===

（2）我将生成的模型生成了在这里

ubuntu@ubuntu:~/tensorrt_inference/arcface/mobilefacenet-res2-6-10-2-dim512$ tree
.
├── mobilefacenet-res2-6-10-2-dim512.caffemodel
├── mobilefacenet-res2-6-10-2-dim512-emore.nchwbin
├── mobilefacenet-res2-6-10-2-dim512-minicaffe.prototxt
├── mobilefacenet-res2-6-10-2-dim512-opencv.prototxt
├── mobilefacenet-res2-6-10-2-dim512.zqparams
├── model-0000.params
├── model-symbol.json
└── onnx├── modelnew2_onnx.onnx└── modelnew_onnx.onnx1 directory, 10 files

（3）然后修改CMakeLists.txt 同样修改一下TensorRT的地址

set(TENSORRT_ROOT /usr/src/tensorrt/)

修改后

set(TENSORRT_ROOT /home/ubuntu/NVIDIA_CUDA-11.1_Samples/TensorRT-7.2.2.3)

对应的配置文件修改

ubuntu@ubuntu:~/tensorrt_inference/arcface/build$ cat ../config.yaml
arcface:onnx_file:     "/home/ubuntu/tensorrt_inference/arcface/mobilefacenet-res2-6-10-2-dim512/onnx/modelnew2_onnx.onnx"engine_file:   "/home/ubuntu/tensorrt_inference/arcface/mobilefacenet-res2-6-10-2-dim512/onnx/arcface_r100.trt"BATCH_SIZE:    1INPUT_CHANNEL: 3IMAGE_WIDTH:   112IMAGE_HEIGHT:  112

（4）编译执行生成tengine文件即可

ubuntu@ubuntu:~/tensorrt_inference/arcface/build$ ./arcface_trt ../config.yaml ../samples/
loading filename from:/home/ubuntu/tensorrt_inference/arcface/mobilefacenet-res2-6-10-2-dim512/onnx/arcface_r100.trt
[11/12/2021-16:31:51] [W] [TRT] TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
deserialize done
[11/12/2021-16:31:51] [W] [TRT] TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
binding0: 150528
binding1: 2048
Processing: ../samples//test9.jpg
prepareImage
prepare image take: 0.246129 ms.
host2device
execute
Inference take: 349.969 ms.
execute success
device2host
post process
Post process take: 0.071033 ms.
Processing: ../samples//test10.jpg
prepareImage
prepare image take: 0.110323 ms.
host2device
execute
Inference take: 0.938029 ms.
execute success
device2host
post process
Post process take: 0.016646 ms.
Processing: ../samples//test7.jpg
prepareImage
prepare image take: 0.085588 ms.
host2device
execute
Inference take: 0.985768 ms.
execute success
device2host
post process
Post process take: 0.036193 ms.
Processing: ../samples//test4.jpg
prepareImage
prepare image take: 0.111465 ms.
host2device
execute
Inference take: 0.919175 ms.
execute success
device2host
post process
Post process take: 0.023314 ms.
Processing: ../samples//test5.jpg
prepareImage
prepare image take: 0.103645 ms.
host2device
execute
Inference take: 0.9132 ms.
execute success
device2host
post process
Post process take: 0.020485 ms.
Processing: ../samples//test3.jpg
prepareImage
prepare image take: 0.104911 ms.
host2device
execute
Inference take: 0.909812 ms.
execute success
device2host
post process
Post process take: 0.022067 ms.
Processing: ../samples//test1.jpg
prepareImage
prepare image take: 0.103754 ms.
host2device
execute
Inference take: 0.91965 ms.
execute success
device2host
post process
Post process take: 0.021067 ms.
Processing: ../samples//test8.jpg
prepareImage
prepare image take: 0.10762 ms.
host2device
execute
Inference take: 0.911278 ms.
execute success
device2host
post process
Post process take: 0.021202 ms.
Processing: ../samples//test6.jpg
prepareImage
prepare image take: 0.085042 ms.
host2device
execute
Inference take: 0.906856 ms.
execute success
device2host
post process
Post process take: 0.014921 ms.
Processing: ../samples//test2.jpg
prepareImage
prepare image take: 0.101656 ms.
host2device
execute
Inference take: 0.918637 ms.
execute success
device2host
post process
Post process take: 0.012409 ms.
Average processing time is 35.9711ms
The similarity matrix of the image folder is:
[1, 0.54835492, 0.53202093, 0.50336587, 0.46709558, 0.52499074, 0.5145005, 0.55176359, 0.53271091, 0.50393772;0.54835492, 1, 0.48129362, 0.4762949, 0.60014594, 0.47247368, 0.52350342, 0.49729002, 0.48303184, 0.53011435;0.53202093, 0.48129362, 1, 0.47364214, 0.49696276, 0.52271658, 0.51096576, 0.48425445, 0.49267167, 0.52210295;0.50336587, 0.4762949, 0.47364214, 1, 0.51976317, 0.80811292, 0.48066908, 0.48449197, 0.45749086, 0.50374758;0.46709558, 0.60014594, 0.49696276, 0.51976317, 1, 0.47129455, 0.53576809, 0.46944529, 0.53187394, 0.5326156;0.52499074, 0.47247368, 0.52271658, 0.80811292, 0.47129455, 1, 0.47609299, 0.48077181, 0.4403888, 0.45222867;0.5145005, 0.52350342, 0.51096576, 0.48066908, 0.53576809, 0.47609299, 1, 0.53024608, 0.54449081, 0.83704132;0.55176359, 0.49729002, 0.48425445, 0.48449197, 0.46944529, 0.48077181, 0.53024608, 1, 0.51537269, 0.48809451;0.53271091, 0.48303184, 0.49267167, 0.45749086, 0.53187394, 0.4403888, 0.54449081, 0.51537269, 1, 0.54323018;0.50393772, 0.53011435, 0.52210295, 0.50374758, 0.5326156, 0.45222867, 0.83704132, 0.48809451, 0.54323018, 1]!

（5）、使用py脚本测试一下

import tensorrt as trt
import pycuda.driver as cuda
import pycuda.autoinit
import numpy as np
import cv2def get_engine(engine_path):# If a serialized engine exists, use it instead of building an engine.print("Reading engine from file {}".format(engine_path))with open(engine_path, "rb") as f, trt.Runtime(TRT_LOGGER) as runtime:return runtime.deserialize_cuda_engine(f.read())TRT_LOGGER = trt.Logger()# engine = get_engine("yolov4_1.trt")def compute_sim(emb1, emb2):from numpy.linalg import normemb1 = emb1.flatten()emb2 = emb2.flatten()sim = np.dot(emb1, emb2) / (norm(emb1) * norm(emb2))return sim# engine = get_engine("mobilefacenet-res2-6-10-2-dim512/onnx/face_reg_mnet.engine")
# print(engine)
# for binding in engine:
#         size = trt.volume(engine.get_binding_shape(binding)) * 1
#         dims = engine.get_binding_shape(binding)
#         print(size)
#         print(dims)
#         print(binding)
#         print(engine.binding_is_input(binding))
#         dtype = trt.nptype(engine.get_binding_dtype(binding))
#         print("dtype = ", dtype)engine = get_engine("/home/ubuntu/tensorrt_inference/arcface/mobilefacenet-res2-6-10-2-dim512/onnx/arcface_r100.trt")
context = engine.create_execution_context()def get_embedding(img):resized = cv2.resize(img, (112, 112), interpolation=cv2.INTER_LINEAR)img_in = cv2.cvtColor(resized, cv2.COLOR_BGR2RGB)img_in = np.transpose(img_in, (2, 0, 1)).astype(np.float32)img_in = np.expand_dims(img_in, axis=0)# img_in /= 255.0img_in = np.ascontiguousarray(img_in)print("Shape of the network input: ", img_in.shape)# print(img_in)# with get_engine("mobilefacenet-res2-6-10-2-dim512/onnx/face_reg_mnet.engine") as engine, engine.create_execution_context() as context:h_input = cuda.pagelocked_empty(trt.volume(context.get_binding_shape(0)), dtype=np.float32)h_output = cuda.pagelocked_empty(trt.volume(context.get_binding_shape(1)), dtype=np.float32)# Allocate device memory for inputs and outputs.d_input = cuda.mem_alloc(h_input.nbytes)d_output = cuda.mem_alloc(h_output.nbytes)# Create a stream in which to copy inputs/outputs and run inference.stream = cuda.Stream()# set the host input datah_input = img_in# print(h_input)# Transfer input data to the GPU.cuda.memcpy_htod_async(d_input, h_input, stream)# Run inference.context.execute_async_v2(bindings=[int(d_input), int(d_output)], stream_handle=stream.handle)# Transfer predictions back from the GPU.cuda.memcpy_dtoh_async(h_output, d_output, stream)# Synchronize the streamstream.synchronize()# Return the host output.# print(h_output)return h_outputimg1 = cv2.imread("/home/ubuntu/tensorrt_inference/arcface/samples/test1.jpg")
emb1 = get_embedding(img1)img2 = cv2.imread("/home/ubuntu/tensorrt_inference/arcface/samples/test2.jpg")
emb2 = get_embedding(img2)print(compute_sim(emb1, emb2))

测试结果

/usr/bin/python3.8 /home/ubuntu/CenterFace/prj-tensorrt/arcface.py
Reading engine from file /home/ubuntu/tensorrt_inference/arcface/mobilefacenet-res2-6-10-2-dim512/onnx/arcface_r100.trt
[TensorRT] WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
[TensorRT] WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
Shape of the network input:  (1, 3, 112, 112)
Shape of the network input:  (1, 3, 112, 112)
0.6905318Process finished with exit code 0

然后整理一个demo进行Centerface和Arface结合进行人脸检测和人证比对，我改了一个简单的代码，在GTX1050显卡测试大约20ms左右

/usr/bin/python3.8 /home/ubuntu/CenterFace/TensorRT_centerface_arcface/demo.py
[TensorRT] WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
[TensorRT] WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
Reading engine from file arcface_r100.trt
[TensorRT] WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
[TensorRT] WARNING: TensorRT was linked against cuBLAS/cuBLAS LT 11.3.0 but loaded cuBLAS/cuBLAS LT 11.2.1
count =  1
face similarity=  0.11867439
gpu times =  26.207 ms
count =  1
face similarity=  0.14468496
gpu times =  21.175 ms
count =  1
face similarity=  0.14031047
gpu times =  20.836 ms
count =  1
face similarity=  0.13255903
gpu times =  20.659 ms
count =  1
face similarity=  0.15876028
gpu times =  20.742 ms

准备移植中；

八、移植Centerface和Arface到Jetson Xavier NX开发板上，

（1）Centerface移植Jetson Xavier Nx

需要将/home/ubuntu/tensorrt_inference 中的Centerface代码中yaml-cpp 依赖全部注释掉，因为开发板上没有这个静态库，我安装过程中发生了错误，索性不安装了，将参数写死就行，生tengine就可以了

我贴一下Centerface中的修改后CMakeLists.txt文件和改后的部分Centerface.cpp文件

cmake_minimum_required(VERSION 3.5)project(CenterFace_trt)set(CMAKE_CXX_STANDARD 14)# CUDA
find_package(CUDA REQUIRED)
message(STATUS "Find CUDA include at ${CUDA_INCLUDE_DIRS}")
message(STATUS "Find CUDA libraries: ${CUDA_LIBRARIES}")# TensorRT
set(TENSORRT_ROOT /usr/src/tensorrt/)
find_path(TENSORRT_INCLUDE_DIR NvInfer.hHINTS ${TENSORRT_ROOT} PATH_SUFFIXES include/)
message(STATUS "Found TensorRT headers at ${TENSORRT_INCLUDE_DIR}")
find_library(TENSORRT_LIBRARY_INFER nvinferHINTS ${TENSORRT_ROOT} ${TENSORRT_BUILD} ${CUDA_TOOLKIT_ROOT_DIR}PATH_SUFFIXES lib lib64 lib/x64)
find_library(TENSORRT_LIBRARY_ONNXPARSER nvonnxparserHINTS  ${TENSORRT_ROOT} ${TENSORRT_BUILD} ${CUDA_TOOLKIT_ROOT_DIR}PATH_SUFFIXES lib lib64 lib/x64)
set(TENSORRT_LIBRARY ${TENSORRT_LIBRARY_INFER} ${TENSORRT_LIBRARY_ONNXPARSER})
message(STATUS "Find TensorRT libs: ${TENSORRT_LIBRARY}")# OpenCV
find_package(OpenCV REQUIRED)
message(STATUS "Find OpenCV include at ${OpenCV_INCLUDE_DIRS}")
message(STATUS "Find OpenCV libraries: ${OpenCV_LIBRARIES}")set(COMMON_INCLUDE ../includes/common)
#set(YAML_INCLUDE ../includes/yaml-cpp/include)#sxj731533730
#set(YAML_LIB_DIR ../includes/yaml-cpp/libs)#sxj731533730#include_directories(${CUDA_INCLUDE_DIRS} ${TENSORRT_INCLUDE_DIR} ${OpenCV_INCLUDE_DIRS} ${COMMON_INCLUDE} ${YAML_INCLUDE})#sxj731533730
include_directories(${CUDA_INCLUDE_DIRS} ${TENSORRT_INCLUDE_DIR} ${OpenCV_INCLUDE_DIRS} ${COMMON_INCLUDE} )
#link_directories(${YAML_LIB_DIR}) #sxj731533730add_executable(CenterFace_trt main.cpp CenterFace.cpp)
#target_link_libraries(CenterFace_trt ${OpenCV_LIBRARIES} ${CUDA_LIBRARIES} ${TENSORRT_LIBRARY} yaml-cpp)  #sxj731533730
target_link_libraries(CenterFace_trt ${OpenCV_LIBRARIES} ${CUDA_LIBRARIES} ${TENSORRT_LIBRARY} )

修改的Centerface.cpp文件

nvidia@nvidia-desktop:~/tensorrt_inference/CenterFace/build$ cat ../CenterFace.cpp
#include "CenterFace.h"
//#include "yaml-cpp/yaml.h"
#include "common.hpp"CenterFace::CenterFace(const std::string &config_file) {//YAML::Node root = YAML::LoadFile(config_file);//YAML::Node config = root["CenterFace"];//YAML::Node root = YAML::LoadFile(config_file);//engine_file = config["engine_file"].as<std::string>();// BATCH_SIZE = config["BATCH_SIZE"].as<int>();// INPUT_CHANNEL = config["INPUT_CHANNEL"].as<int>();// IMAGE_WIDTH = config["IMAGE_WIDTH"].as<int>();// IMAGE_HEIGHT = config["IMAGE_HEIGHT"].as<int>();onnx_file = "../centerface_480_640.onnx";engine_file ="../centerface.trt";BATCH_SIZE = 1;INPUT_CHANNEL = 3;IMAGE_WIDTH = 640;IMAGE_HEIGHT = 480;obj_threshold = 0.5;nms_threshold = 0.45;
}

在Jetson Xavier Nx执行即可

nvidia@nvidia-desktop:~/tensorrt_inference/CenterFace/build$ make
Scanning dependencies of target CenterFace_trt
[ 33%] Building CXX object CMakeFiles/CenterFace_trt.dir/main.cpp.o
[ 66%] Building CXX object CMakeFiles/CenterFace_trt.dir/CenterFace.cpp.o
[100%] Linking CXX executable CenterFace_trt
[100%] Built target CenterFace_trt
nvidia@nvidia-desktop:~/tensorrt_inference/CenterFace/build$ ./CenterFace_trt ../config.yaml  ../samples/
/config.yaml  ../samples/
loading filename from:../centerface.trt
deserialize done
binding0: 3686400
binding1: 76800
binding2: 153600
binding3: 153600
binding4: 768000
Processing: ../samples//test2.jpg

（2）Arface也是类似处理,不详细叙述

vidia@nvidia-desktop:~/tensorrt_inference/arcface/build$ make
Scanning dependencies of target arcface_trt
[ 33%] Building CXX object CMakeFiles/arcface_trt.dir/main.cpp.o
[ 66%] Building CXX object CMakeFiles/arcface_trt.dir/arcface.cpp.o
[100%] Linking CXX executable arcface_trt
[100%] Built target arcface_trt
nvidia@nvidia-desktop:~/tensorrt_inference/arcface/build$ ./arcface_trt ../config.yaml ../samples/
----------------------------------------------------------------
Input filename:   ../modelnew2_onnx.onnx
ONNX IR version:  0.0.8
Opset version:    15
Producer name:
Producer version:
Domain:
Model version:    0
Doc string:
----------------------------------------------------------------
start building engine

最终测试一下开发板的人脸检测、比对效果蛮好的 PC端GTX1050最快达到30fps,

Jetson Xavier NX开发板最快达到15fps 很流畅

Jetson Xavier NX开发板

开发板的显存使用率和功耗都不是很大 NVIDIA yyds

整个代码上传GitHub： https://github.com/sxj731533730/TensorRT_centerface_arcface

里面含有两个文件夹模型一个模型针对PC 的cuda11.2+TensorRT7.2.2 另一个文件夹模型针对Jetson Nano NX 的cuda10.2+TensorRT 7

八、突然想到使用Mediapipe进行人脸的关键点检测和比对是否效果是不是也会不错，发现的确还挺好的，即使使用tensorflow-cpu也达到了14~20FPS 但是在开发板上，太吃cpu了，而且GPU的Tensorflow版本没搭建起来，有空再说

九、开始做串口数据转发，对接STM32

首先购买了京东商店的转接模块六合一串口模块 - 丢石头百科测试，下载驱动，安装驱动，然后测试

测试一下，该usb转ttl可用

测试

sudo chmod 777 /dev/ttyTHS1

测试代码

import cv2
import serial
import struct,time
import time
import binascii
num=[None]*5
ser=serial.Serial("/dev/ttyTHS1",115200) #
def recv():print("receive test.......")while True:for i in range (0,5):data=str(binascii.b2a_hex(ser.read(1)))[2:-1]num[i]=dataif((num[0]=='5a') and (num[1]=='a5')):print(num)
def write():print("write test.......")while True:ser.write("666".encode("UTF-8"))print("36 36 36")time.sleep(1)
op=input("enter the operation:")
if op =="0":recv()
elif op=="1":write()

硬件图

URAT的引针第三排内侧为RTX 外侧为TXD 第三排最后一个外侧为GND （外侧指边缘）

连线图

橙色接线连接USB转TTL开发板的 TXD 连接开发版的RXD引针

红色接线连接USB转TTL开发板的 RXD 连接开发版的TXD引针

绿色接线连接USB转TTL开发板的 GND 连接开发版的GND引针

其实没必要连接USB转TTL开发板的 TXD 又不从PC端发数据给开发板

测试图

参考

通过VNC搭建Ubuntu 18.04和20.04图形界面 - 轻量应用服务器 - 阿里云

CenterFace+TensorRT部署人脸和关键点检测400fps - 知乎

笔记（四）Jetson Nano 系统登录_SWORLD-CSDN博客_jetson nano密码

Install PyTorch on Jetson Nano - Q-engineering

GitHub - zheshipinyinMc/arcface_retinaface_mxnet2onnx: arcface and retinaface model convert mxnet to onnx.

21、Jetson Xavier Nx使用Centerface+Arcface进行人脸识别、认证、比对。并且使用USB转TTL进行串口通信相关推荐

NVIDIA Jetson Xavier NX设备上使用jtop监控GPU、CPU、内存等的使用
文章目录: 1 在Jetson设别上安装jtop 2 jtop的使用 2.1 jtop的使用 2.2 jetson-stats一些可执行文件使用 3 jetson NX默认刷好机之后的环境 1 在Je ...
在Jetson Xavier NX上安装pycuda报错：src/cpp/cuda.hpp:14:10: fatal error: cuda.h: No such file or directory
文章目录: 1 我的系统环境和遇到问题分析 1.1 我的系统环境 1.2 问题描述 2 问题解决方式 1 我的系统环境和遇到问题分析 1.1 我的系统环境我的详细系统环境如下:使用jetson_re ...
在NVIDIA Jetson Xavier NX上把yolov4-deepsort的模型pb模型使用tensorflow-onnx和onnx-tensorrt工具最终转换为tensorrt模型
文章目录: 1 安装tensorflow-onnx环境和把tensorflow的pb模型转换为onnx模型 1.1 安装tensorflow2onnx环境 1.2 把tensorflow的pb模型转换 ...
NVIDIA Jetson Xavier NX上导入tensorflow报错:AttributeError: module ‘wrapt‘ has no attribute ‘ObjectProxy‘
欢迎大家关注笔者,你的关注是我持续更博的最大动力原创文章,转载告知,盗版必究在Jetson Xavier NX上导入tensorflow报错:AttributeError: module 'wra ...
nvidia的jetson系列的方案_NVIDIA Jetson Xavier NX开发者套件主要应用于自主机器边缘计算产品系列...
NVIDIA副总裁,边缘计算业务总经理Deepu Talla表示:"NVIDIA已在多个行业中创建并优化了大量AI软件.如今,有了云原生技术的加持,我们在嵌入式和机器人领域的客户就能通过高性 ...
仅信用卡大小，性能比TX2强15倍，英伟达边缘AI计算新品Jetson Xavier NX
点击我爱计算机视觉标星,更快获取CVML新技术晓查发自凹非寺量子位报道 | 公众号 QbitAI 英伟达最近发布了Jetson Xavier NX,这是一个用于在无人机.汽车和机器人等边缘 ...
Jetson Xavier NX （11） -- NX介绍与系统烧录
目录 1. Jetson Xavier NX介绍 1.1 NX 性能 1.2 硬件总览 1.3 相关资料 2. 系统烧录 2.1 下载官方镜像 2.2 格式化SD卡 2.3 烧录系统 3 测试 1 ...
胎儿式保姆级教程：Jetson Xavier NX镜像烧录、开机配置、中文配置、风扇设置、远程桌面、文件传输配置、pycharm安装环境配置,QQ,opencv(cuda编译),torch(GPU).
一.使用的硬件:nx板子.原装充电器.有线鼠标(USB).有线键盘(USB).有线摄像头(USB).7寸触摸显示屏.20寸大显示屏(显示屏有一个就可以,大屏幕更加方便).SD卡(128G,用64g也可 ...
JETSON XAVIER NX 入门教程（一）入手刷机过程
转载自:https://mp.weixin.qq.com/s/9HFZaP3A4cslJpuU5AqmGQ JETSON XAVIER NX 入门教程(一)入手刷机过程原创 CXY 北理玄驹合作小组 ...
【模型部署】Jetson Xavier NX（eMMC）部署YOLOv5-5.0
文章目录前言 NVIDIA Jetson Jetson Xavier NX 版本区别(SD | eMMC) 规格参数 Jetpack4.6.1环境搭建烧录系统(OS) SSD启动 SSD分区设置 ...

21、Jetson Xavier Nx使用Centerface+Arcface进行人脸识别、认证、比对。并且使用USB转TTL进行串口通信

21、Jetson Xavier Nx使用Centerface+Arcface进行人脸识别、认证、比对。并且使用USB转TTL进行串口通信相关推荐

最新文章

热门文章