地平线bev参考算法的官方文档中介绍了单帧可视化推理脚本infer.py的使用，输入为6V环视图像和homography矩阵，6V环视图像可以从解压的nuscenes数据集中获取，homography矩阵则是需要从nuscenes数据集中获取相机参数信息后经计算获得。
本文将介绍一种简单的homography矩阵导出方法，输入为nuscens数据集中6张环视图片的文件名"fileneme"，输出为homography矩阵。
1. 安装nuscenes-devkit
首先在python环境中安装nuscenes-devkit，参考命令如下 ：
pip install nuscenes-devkit
2. v1.0-mini数据集下载
在nuscenes官网（https://www.nuscenes.org/nuscenes#download）下载v1.0-mini数据集后解压，解压后的目录如下所示：
├── v1.0-mini
├── maps
├── samples
├── sweeps
└── v1.0-mini
3. 导出homography矩阵的代码
from nuscenes.nuscenes import NuScenes
import numpy as np
from pyquaternion import Quaternion
def get_homography_by_cam(sensor2ego_translation, sensor2ego_rotation, camera_intrinsic):
#将旋转矩阵转为四元数
rotation = Quaternion(sensor2ego_rotation).rotation_matrix
ego2sensor_r = np.linalg.inv(rotation)
ego2sensor_t = sensor2ego_translation @ ego2sensor_r.T
ego2sensor = np.eye(4)
ego2sensor[:3, :3] = ego2sensor_r.T
ego2sensor[3, :3] = -np.array(ego2sensor_t)
camera_intrinsic = np.array(camera_intrinsic)
viewpad = np.eye(4)
viewpad[
: camera_intrinsic.shape[0], : camera_intrinsic.shape[1]
] = camera_intrinsic
ego2img = viewpad @ ego2sensor.T
return ego2img
if __name__ == "__main__":
#初始化，这里以v1.0-mini数据集为例
nusc = NuScenes(version='v1.0-mini', dataroot='../v1.0-mini', verbose=True)
#6张图像的文件名
#要注意这里的输入顺序必须遵循：
#FRONT_LEFT,FRONT,FRONT_RIGHT,BACK_LEFT,BACK,BACK_RIGHT
image_list=[
'n008-2018-08-01-15-16-36-0400__CAM_FRONT_LEFT__1533151603504799.jpg',
'n008-2018-08-01-15-16-36-0400__CAM_FRONT__1533151603512404.jpg',
'n008-2018-08-01-15-16-36-0400__CAM_FRONT_RIGHT__1533151603520482.jpg',
'n008-2018-08-01-15-16-36-0400__CAM_BACK_LEFT__1533151603547405.jpg',
'n008-2018-08-01-15-16-36-0400__CAM_BACK__1533151603537558.jpg',
'n008-2018-08-01-15-16-36-0400__CAM_BACK_RIGHT__1533151603528113.jpg'
]
# 获取数据集中所有的sample
# 包含'token','calibrated_sensor_token','filename','channel'等信息
sample_data=nusc.sample_data
# 获取传感器定义：
# 包含'token','sensor_toker','translation','rotation','camera_intrinsic'等信息
cali_sensor=nusc.calibrated_sensor
translation=[]
rotation=[]
cam_intrinsic=[]
ego2imgs=[]
for i,img in enumerate(image_list):
print(img)
for sample in sample_data:
if img in sample["filename"]:
#1. 选择输入图片的name来获取捕获该图片的"calibrated_sensor_token"
calibrated_sensor_token=(sample["calibrated_sensor_token"])
print("calibrated_sensor_token:",calibrated_sensor_token)
#2.根据"calibrated_sensor_token"获取传感器的"translation"、
# "rotation"、和"camera_intrinsic"等参数
for sensor in cali_sensor:
if sensor["token"]==calibrated_sensor_token:
#获取translation矩阵
sensor2ego_translation=np.array(sensor["translation"])
translation.append(sensor2ego_translation)
#获取rotation矩阵
sensor2ego_rotation=np.array(sensor["rotation"])
rotation.append(sensor2ego_rotation)
#获取相机内参矩阵
camera_intrinsic=np.array(sensor["camera_intrinsic"])
cam_intrinsic.append(camera_intrinsic)
#计算homography矩阵
ego2img=get_homography_by_cam(
sensor2ego_translation,
sensor2ego_rotation,
camera_intrinsic
)
ego2imgs.append(ego2img)
#导出translation矩阵，旋转矩阵、相机内参矩阵、homo矩阵
translation=np.array(translation)
np.save("translation",translation)
rotation=np.array(rotation)
np.save("rotation",rotation)
cam_intrinsic=np.array(cam_intrinsic)
np.save("cam_intrinsic",cam_intrinsic)
ego2imgs=np.array(ego2imgs)
np.save("ego2imgs",ego2imgs)
生成的ego2imgs.npy即为homography矩阵。