Tracknet usage record: environment configuration

First of all , The main configuration requirements come from the description of the open source website of the project , Here, record the pit you stepped on
Project to connect TrackNetV2

New virtual environment

The installation environment of the original author is Ubuntu+Python3.5.2+CUDA10.1, But my computer is win11+Anaconda Environmental Science , The configuration is similar to that of the author .
First, create a corresponding version python A virtual environment . Note that the version corresponds to .python Version is 3.5.2
By default, you have installed Anaconda And be able to cmd Use in
Create directly on the command line ：

coda create -n tracknet-gpu python=3.5.2

install CUDA

If you want to use it GPU Reasoning （ Be sure to use GPU,CPU The reasoning speed of is very, very slow ）, Then install related package Before you do that, you have to CUDA It's installed correctly .
Due to the use of source code tensorflow To build the network , Needed CUDA The version should correspond to the .
Here is the first important point ： The version used by the author is TensorFlow 1.13.1/keras 2.2.4/Opencv 4.1.0/CUDA 10.1 But this configuration is win Measured in the environment cannot be used ！
according to Tensorflow The version test data of the official website includes ：


We're going to use tensorflow_gpu-1.13.1, Corresponding CUDA Version is 10.0,cuDNN Version is 7.4
About CUDA Installation has been described in detail in many blogs , Please search for . About installation , It is best to use when the computer has enough storage CUDA Multiple versions coexist , It is convenient for later version switching development .

install Package

The installation command given by the author ：

  $ sudo apt-get install git
  $ sudo apt-get install python3-pip
  $ pip3 install pyqt5
  $ pip3 install pandas
  $ pip3 install PyMySQL
  $ pip3 install opencv-python
  $ pip3 install imutils
  $ pip3 install Pillow
  $ pip3 install piexif
  $ pip3 install -U scikit-learn
  $ pip3 install keras
  $ git clone https://nol.cs.nctu.edu.tw:234/open-source/TrackNetv2

We need to change the adaptation here cmd operation . Created before entering tracknet-gpu After environment ：

coda install git
pip install pyqt5
pip install tensorflow-gpu==1.13.1
pip install pandas
pip install PyMySQL
pip install opencv-python==4.1.0
pip install imutils
pip install Pillow
pip install piexif
pip install scikit-learn
pip install keras

If you don't surf the Internet scientifically , These packages cannot all be installed .
Curve of national salvation ： stay pypi Official website Download the required package , Then on the command line cd Download path to package , And then again pip install You can install locally .
I downloaded the following packages ：

Refer to the package name in the picture 、 edition 、 Download the compilation platform to the local , And then install it
It is worth mentioning that ：Keras Version and tensorflow The version of is also corresponding .
During the installation process, because python3.5 This version has stopped supporting ,pip There will be warnings during installation , Can be ignored .
The installation process is based on , Put the package given by the author under a good installation condition , One by one installation , Search for any dependent packages you lack 、 download 、 Local installation , Lack of Han make up what . Until all packages are installed .

testing

After installing the environment correctly , Clone the whole project on the project website , Looking for a video to be detected . Test command

python predict.py --video_name=<videoPath> --load_weights=<weightPath>

among videoPath Is the video path to be detected ,weightPath There are two choices ：model_33 and model906_30, The previous model is three in one detection , The latter model is three frame three check . What is the difference between the two tests? You can read the author's paper .
The network structure of three in one detection is like this ,9 Dimension input ,1 Dimension output .

The output result after completing the test is CSV file , It contains the frame time point of the frame that recognizes the ball and the coordinate point of the ball .
Overlap the detection data with the original video , You can get the video that marks the ball movement .

python show_trajectory.py <input_video_path> <input_csv_path>

input_video_path Is the video path detected in the previous step ,input_csv_path It is generated after the previous detection csv File path .