Terrain classification

Author: Shiwei Liu, College of Artificial Intelligence, Nankai University

E-mail: [email protected]

About this program

This project is my final assignment for the course "Deep learning" during my junior year(2022) at Nankai University.It is used to train a terrain classification model in order to distinguish where the robot is running thus we can switch the controllling policy of it. I used AlexNet for reference and built a simular net structure despite the numbers of GPU (My PC only has one GPU RTX 3060). After training for 50 epochs, it can achieve 96.575% of train accuracy and 96.125% of test accuracy.

Requirements

1. You need a conda enviroment with a suitable Python version:

   conda create -n YOUR_ENV_NAME python==3.9
   conda activate YOUR_ENV_NAME
   

2. Install pytorch(GPU version with your CUDA version), matplotlib, tqdm and numpy:

   pip3 install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu117
   pip install matplotlib
   pip install tqdm
   pip install numpy
   

Run this program

1. I provide this terrian dataset in the project file"dixing2" which is pre-seperated by data_seperate.py. If you wants to use the source dataset in your way here it is.
2. To train this model, run:

   python Building.py -- run_mode train --train_epochs YOUR_TRAIN_EPOCHS
   

   To test the result, run:

   python Building.py -- run_mode test
   

Result

This dataset is plenty easy for this network. I tested this model on 800 images and got a 96.125% of accuracy after only 50 training epochs.

Test Figure 2 is taken while the robot is moving at high speed, and it also get a outstanding prediction.

I also show the training and evalidation loss and accuracy curve here: