This project is about how a Realsense L515 can detect the Pipe Type which is left right or open in real-time at the ROS platform. The point cloud data is captured by Realsense L515 and Gemini camera. The models were trained using PointNet++ method.
The results show that using the SSG method can reach 95.83% accuracy in predicting the pipe type which is faster than the MSG method (91.67%). It proves that the SSG method is good at predicting pipe type.
The models were trained at the windows side using Anaconda. The configurations and settings are shown below:
| Dependencies | VERSION |
|---|---|
| CUDA | 12.1 |
| CUDNN | 8.9.7.29 |
| Python | 3.8 |
| PyTorch | 2.1.0 |
| PyTorch Cuda | 12.1.0 |
| Torch Vision | 0.16.0 |
Laptop configuration:
| DETAIL | VERSION |
|---|---|
| Processer | Intel i7-13700H |
| Ram | 16 GB |
| NVIDIA driver | 536.45 |
| GPU | RTX4060-GPU 8 GB |
Code below is used for setting up environment in conda:
conda activate pointnet2 #pointnet2 is my environment name in conda
conda install pytorch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 pytorch-cuda12.1 -c pytorch -c nvidia -y