文章标题:SeqTrack3D: Exploring Sequence Information for Robust 3D Point Cloud Tracking

SeqTrack3D 是东北大学团队提出的 面向 3D 点云场景的长时跟踪算法,核心突破是打破传统方法对 “单帧 / 相邻两帧” 的依赖,通过 显式建模多帧点云的时空序列关联,解决 3D 点云跟踪中稀疏性、目标遮挡、快速运动等核心痛点,在自动驾驶、机器人导航等场景中具备强实用性。

1. 环境安装

创建环境

conda create -n seqtrack3d python=3.9 -y

激活环境

conda activate seqtrack3d

安装torch, torchvision, torchaudio

/root/miniconda3/envs/seqtrack3d/bin/pip install torch==2.1.0 torchvision==0.16.0 torchaudio==2.1.0 --index-url https://download.pytorch.org/whl/cu118

安装numpy

/root/miniconda3/envs/seqtrack3d/bin/pip install numpy==1.23.5

安装protobuf, easydict, pandas

/root/miniconda3/envs/seqtrack3d/bin/pip install protobuf easydict pandas

下载Pointnet2_PyTorch

https://github.com/erikwijmans/Pointnet2_PyTorch.git

进入Pointnet2_PyTorch

cd Pointnet2_PyTorch/pointnet2_ops_lib

安装Pointnet2_PyTorch

/root/miniconda3/envs/seqtrack3d/bin/pip install . --no-build-isolation

安装pomegranate

/root/miniconda3/envs/seqtrack3d/bin/pip install pomegranate==0.14.8 --no-build-isolation

安装其它依赖

/root/miniconda3/envs/seqtrack3d/bin/pip install pyquaternion pytorch-lightning PyYAML Shapely tdqm tensorboard torchmetrics nuscenes-devkit

2. 训练

python3 main.py --cfg cfgs/seqtrack3d_nuscenes.yaml --batch_size 4 --epoch 20 --seed 42 --tag "v0.0.1"

参考文献

https://github.com/aron-lin/seqtrack3d?tab=readme-ov-file

Logo

鲲鹏昇腾开发者社区是面向全社会开放的“联接全球计算开发者,聚合华为+生态”的社区,内容涵盖鲲鹏、昇腾资源,帮助开发者快速获取所需的知识、经验、软件、工具、算力,支撑开发者易学、好用、成功,成为核心开发者。

更多推荐