This repository contains source code and models for BEVFusion online real-time inference using CUDA, TensorRT & ROS.
Supports ROS2. please switch to the galactic-devel branch
ubuntu-20.04,noetic,cuda-11.3, cudnn-8.6.0, TensorRT-8.5
-
默认已安装
noetic, cuda, cudnn, 已下载TensorRT源码 -
ros依赖
# 1. 建立ros工作空间
mkdir -p bevfusion_ws/src
# 2. 进入bevfusion_ws/src目录，拉取源码
cd bevfusion_ws/src
git clone https://github.com/linClubs/BEVFusion-ROS-TensorRT.git
# 3. 进入bevfusion_ws工作空间一键安装功能包需要ros依赖
cd ..
rosdep install -r -y --from-paths src --ignore-src --rosdistro $ROS_DISTRO- 修改
./tool/environment.sh中cuda tensorrt cudnn的路径, 运行./tool/build_trt_engine.sh生成tensorrt推理模型
./tool/build_trt_engine.shros包准备
-
bevfusion官方提供了已训练好的nuscenes模型参数 -
rosbag数据转换参考
nuscenes2rosbag功能包 -
nuscenes传感器之间的参数已给出,无需标定
如果需接真实的传感器进行场景测试,需提前完成训练和标定工作
- 编译前需要修改
CMakeLists.txt中TensorRT和CUDA路径,修改如下
...
# cuda
set(CUDA_TOOLKIT_ROOT_DIR /usr/local/cuda-11.3) # CUDA修改这一行
set(CUDA_INSTALL_TARGET_DIR targets/x86_64-linux)
set(CUDA_INCLUDE_DIRS ${CUDA_TOOLKIT_ROOT_DIR}/${CUDA_INSTALL_TARGET_DIR}/include)
set(CUDA_LIBS ${CUDA_TOOLKIT_ROOT_DIR}/${CUDA_INSTALL_TARGET_DIR}/lib)
# TENSORRT
set(TensorRT_ROOT /home/lin/software/TensorRT-8.5.3.1) # TensorRT修改这一行
# set(TensorRT_ROOT ~/share/TensorRT-8.5.3.1)
set(TensorRT_INCLUDE_DIRS ${TensorRT_ROOT}/include)
set(TensorRT_LIBS ${TensorRT_ROOT}/lib/)
...- 编译运行
bevfusion_node.launch修改model_name与precision参数值
model_name: resnet50/resnet50int8/swint
precision: fp16/int8
swint + int8模式不能工作
# 1. 编译
catkin_make
# 2. source工作空间
source devel/setup.bash
# 3. 运行bevfusion_node
roslaunch bevfusion bevfusion_node.launch
# 4. 播放数据集
rosbag play 103.bag - 运行报错
tool/simhei.ttf找不到, 全局搜索tool/simhei.ttf或者UseFont关键字
在/src/common/visualize.cu中修改UseFont的值即可,改成simhei.ttf正确的路径即可
-
bev感知交流群-472648720, 欢迎各位小伙伴进群一起学习讨论bev相关知识！！！^_^
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent