Azure Kinect DK相机单帧扫描的图像会存在一定的信息失真和缺失,于是想做个多帧重建试试效果。
非实时重建
这里主要参考这篇文章:《Azure Kinect DK 实现三维重建 (PC非实时版)》。
博主 lucky li 给的代码可以跑通,效果也算可以。整个代码的流程大致是录制一段重建对象的视频,然后进行逐帧分解得到彩色图像和深度图像,然后每帧相互之间进行特征匹配,最终重建点云。这里对博主的代码进行了一定的调整,因为录制视频的代码可视化窗口只显示彩色图像和深度图像,Kinect相机在距离过近时采集信息会失真或缺失,采用显示点云便于判断相机距离是否合适。
首先将以下GitHub中的pykinect_azure文件夹复制到博主代码的open3d_reconstruction/sensor文件夹下:https://github.com/struggler176393/Kinect_pointcloud
open3d_reconstruction/sensors/azure_kinect_recorder.py修改如下:
import sys
sys.path.insert(1, './')
import argparse
import datetime
import open3d as o3d
import numpy as np
import pykinect_azure as pykinect
from pykinect_azure.utils import Open3dVisualizer
class RecorderWithCallback:
def __init__(self, config, device, filename, align_depth_to_color):
# Global flags
self.flag_exit = False
self.flag_record = False
self.filename = filename
self.open3dVisualizer = Open3dVisualizer()
self.align_depth_to_color = align_depth_to_color
self.recorder = o3d.io.AzureKinectRecorder(config, device)
if not self.recorder.init_sensor():
raise RuntimeError('Failed to connect to sensor')
def escape_callback(self, vis):
self.flag_exit = True
if self.recorder.is_record_created():
print('Recording finished.')
else:
print('Nothing has been recorded.')
return False
def space_callback(self, vis):
if self.flag_record:
print('Recording paused. '
'Press [Space] to continue. '
'Press [ESC] to save and exit.')
self.flag_record = False
elif not self.recorder.is_record_created():
if self.recorder.open_record(self.filename):
print('Recording started. '
'Press [SPACE] to pause. '
'Press [ESC] to save and exit.')
self.flag_record = True
else:
print('Recording resumed, video may be discontinuous. '
'Press [SPACE] to pause. '
'Press [ESC] to save and exit.')
self.flag_record = True
return False
def run(self):
glfw_key_escape = 256
glfw_key_space = 32
self.open3dVisualizer.vis.register_key_callback(glfw_key_escape, self.escape_callback)
self.open3dVisualizer.vis.register_key_callback(glfw_key_space, self.space_callback)
print("Recorder initialized. Press [SPACE] to start. "
"Press [ESC] to save and exit.")
vis_geometry_added = False
while not self.flag_exit:
self.rgbd = self.recorder.record_frame(self.flag_record,
self.align_depth_to_color)
if self.rgbd is None:
continue
color_image = self.rgbd.color
depth_image = self.rgbd.depth
rgbd_image = o3d.geometry.RGBDImage.create_from_color_and_depth(
color_image, depth_image,
convert_rgb_to_intensity=False)
intrinsic = o3d.camera.PinholeCameraIntrinsic(
o3d.camera.PinholeCameraIntrinsicParameters.PrimeSenseDefault )
intrinsic.set_intrinsics(
width=1280, height=720, fx=605.805115, fy=605.625549, cx=641.717163, cy=363.225800)
self.point_cloud = o3d.geometry.PointCloud.create_from_rgbd_image(rgbd_image, intrinsic)
self.open3dVisualizer(self.point_cloud.points,self.point_cloud.colors)
self.recorder.close_record()
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Azure kinect mkv recorder.')
parser.add_argument('--config', type=str, help='input json kinect config')
parser.add_argument('--output', type=str, help='output mkv filename')
parser.add_argument('--list',
action='store_true',
help='list available azure kinect sensors')
parser.add_argument('--device',
type=int,
default=0,
help='input kinect device id')
parser.add_argument('-a',
'--align_depth_to_color',
action='store_false',
help='enable align depth image to color')
args = parser.parse_args()
if args.list:
o3d.io.AzureKinectSensor.list_devices()
exit()
if args.config is not None:
config = o3d.io.read_azure_kinect_sensor_config(args.config)
else:
config = o3d.io.AzureKinectSensorConfig()
if args.output is not None:
filename = args.output
else:
filename = '{date:%Y-%m-%d-%H-%M-%S}.mkv'.format(
date=datetime.datetime.now())
print('Prepare writing to {}'.format(filename))
device = args.device
if device < 0 or device > 255:
print('Unsupported device id, fall back to 0')
device = 0
r = RecorderWithCallback(config, device, filename,
args.align_depth_to_color)
r.run()
由于我要重建的对象尺寸比较小,因此需要修改一下运行python sensors/azure_kinect_mkv_reader.py --input dataset/name.mkv --output dataset/name
之后得到的open3d_reconstruction/dataset/name/config.json文件,减小tsdf_cubic_size
参数的值(找半天才找到),最终得到的scene/integrated.ply文件点云稠密性越高,尝试了一下tsdf_cubic_size
取0.5和0.3看看效果,取0.2及以下的时候电脑跑不动杀死进程了。
一些运行的时间数据(录制了7秒的视频):
- Making fragments 0:06:34.716390
- Register fragments 0:00:00.016120
- Refine registration 0:00:00.015653
tsdf_cubic_size=0.5
Integrate frames 0:00:56.776975
点云结果:
tsdf_cubic_size=0.3
Integrate frames 0:05:38.702845
点云结果:
- 两种参数得到的结果基本相差不大,但是重建还是比较耗时的,难以满足实时性的需求,由于这次重建大概有71帧RGB图像和深度图像参与重建,如果想要加快速度可以减小帧数试试。还有就是多帧重建实际上好像牺牲了一定的精度,下面这是单帧扫描的点云图像:
单帧点云图除了有些地方可能缺失深度信息,精度方面还是更好一些,多帧重建应该更适合大环境重建。
实时重建
这里主要参考这篇文章:《Azure Kinect DK 实现三维重建 (jetson实时版)》。
ROS工作空间建好编译就可以看结果了:
基本上大致的物体是能重建出来的,就是不知道为什么有很多从相机位置发散出来的噪声点,导致点云有点不好看,后续可以去噪处理一下。