fine2copyv / bevfusion-ros-tensorrt

BEVFusion-ROS-TensorRT

BEVFusion-ROS-TensorRT-CPP

This repository contains source code and models for BEVFusion online real-time inference using CUDA, TensorRT & ROS.

1 依赖安装

• ubuntu-20.04,noetic,cuda-11.3, cudnn-8.6.0, TensorRT-8.5

1. 默认已安装noetic, cuda, cudnn, 已下载TensorRT源码

2. 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

3. 模型下载参考

4. 模型导出参考

5. ros包准备

• bevfusion官方提供了已训练好的nuscenes模型参数
• 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

传感器标定参考

nuscenes2rosbag

2 编译运行

1. 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/)
...

2. 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 

---

BEVFusion-ROS-TensorRT部分函数介绍

1. 点云float32转float16

"points.tensor" 是 float16 数据格式. c++浮点数一般用float4个字节

typedef unsigned short half;
static inline half __internal_float2half(const float f);

// msg转half精度在换成nv::DataType::Float16
void cloud_cb(const sensor_msgs::PointCloud2::ConstPtr &msg)
{
    pcl::PointCloud<pcl::PointXYZI>::Ptr ROI_cloud(new pcl::PointCloud<pcl::PointXYZI>);
    // msg转成pcl
    pcl::fromROSMsg(*msg, *ROI_cloud);

    std::cout << "ROI_cloud->points.size()   " << ROI_cloud->points.size() << std::endl;
    
    half *points = new half[ROI_cloud->points.size() * 5];
    
    for (int i = 0; i < ROI_cloud->points.size(); i++)
    {
        points[i * 5 + 0] = __internal_float2half(ROI_cloud->points[i].x);
        points[i * 5 + 1] = __internal_float2half(ROI_cloud->points[i].y);
        points[i * 5 + 2] = __internal_float2half(ROI_cloud->points[i].z);
        points[i * 5 + 3] = __internal_float2half(1);
        points[i * 5 + 4] = __internal_float2half(0);    // 实时激光雷达没有第五维数据可直接赋值0
    }
    vector<int32_t> shape{ROI_cloud->points.size(), 5};

    // Tensor Tensor::from_data_reference(void *data, vector<int32_t> shape, DataType dtype, bool device)

    // float32转float16
    lidar_point_cloud_tensor = nv::Tensor::from_data_reference(points, shape, nv::DataType::Float16, false);
}

2. opencv读取图像转stb库支持的格式

unsigned char* cv2stb(std::string img_path)
{
  cv::Mat image = cv::imread(img_path.c_str(), cv::IMREAD_UNCHANGED);
  int width = image.cols;  // 宽x
  int height = image.rows; // 高y
  int channels = image.channels(); // 通道

  std::vector<unsigned char> buffer;    // 创建一个char类型的数组buffer用来存储图像的data域
  cv::imencode(".jpg", image, buffer); // 编码格式 ""参数可添 .jpg、.png

  // 使用stbi_load函数加载图像数据 width * height * channels = buffer.size()
  unsigned char* stbi_data = stbi_load_from_memory(buffer.data(), buffer.size(), &width, &height, &channels, 0);
  return stbi_data;
}

3. 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

---

References

• bevfusion

• Lidar_AI_Solution

• 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！！！^_^