双目立体视觉- ZED2 ROS 双目视觉开发理论与实践 2

Posted 2021-11-06 Techblog of HaoWANG

ROS - ZED Node

To start a ZED ROS node you can use the following commands in a shell console:

• ZED:

$ roslaunch zed_wrapper zed.launch

• ZED Mini:

$ roslaunch zed_wrapper zedm.launch

• ZED 2:

$ roslaunch zed_wrapper zed2.launch

• ZED 2i:

$ roslaunch zed_wrapper zed2i.launch

Published Topics

The ZED node publishes data to the following topics:

• Left camera

  • rgb/image_rect_color: Color rectified image (left RGB image by default)
  • rgb/image_rect_gray: Grayscale rectified image (left RGB image by default)
  • rgb_raw/image_raw_color: Color unrectified image (left RGB image by default)
  • rgb_raw/image_raw_gray: Grayscale unrectified image (left RGB image by default)
  • rgb/camera_info: Color camera calibration data
  • rgb_raw/camera_info: Color unrectified camera calibration data
  • left/image_rect_color: Left camera color rectified image
  • left/image_rect_gray: Left camera grayscale rectified image
  • left_raw/image_raw_color: Left camera color unrectified image
  • left_raw/image_raw_gray: Left camera grayscale unrectified image
  • left/camera_info: Left camera calibration data
  • left_raw/camera_info: Left unrectified camera calibration data

• Right camera

  • right/image_rect_color: Color rectified right image
  • right_raw/image_raw_color: Color unrectified right image
  • right/image_rect_gray: Grayscale rectified right image
  • right_raw/image_raw_gray: Grayscale unrectified right image
  • right/camera_info: Right camera calibration data
  • right_raw/camera_info: Right unrectified camera calibration data

• Stereo pair

  • stereo/image_rect_color: stereo rectified pair images side-by-side
  • stereo_raw/image_raw_color: stereo unrectified pair images side-by-side

  Note: to retrieve the camera parameters you can subscribe to the topics left/camera_info, right/camera_info, left_raw/camera_info and right_raw/camera_info

• Depth and point cloud

  • depth/depth_registered: Depth map image registered on left image (32-bit float in meters by default)
  • depth/camera_info: Depth camera calibration data
  • point_cloud/cloud_registered: Registered color point cloud
  • confidence/confidence_map: Confidence image (floating point values to be used in your own algorithms)
  • disparity/disparity_image: Disparity image

• Tracking

  • odom: Absolute 3D position and orientation relative to the Odometry frame (pure visual odometry for ZED, visual-inertial for ZED-M and ZED 2)
  • pose: Absolute 3D position and orientation relative to the Map frame (Sensor Fusion algorithm + SLAM + Loop closure)
  • pose_with_covariance: Camera pose referred to Map frame with covariance
  • path_odom: Sequence of camera odometry poses in Map frame
  • path_map: Sequence of camera poses in Map frame

• Mapping

  • mapping/fused_cloud: Fused color point cloud

  Note: published only if mapping is enabled, see mapping/mapping_enabled parameter

• Sensor Data

  • imu/data: Accelerometer, gyroscope, and orientation data in Earth frame [only ZED-M and ZED 2]
  • imu/data_raw: Accelerometer and gyroscope data in Earth frame [only ZED-M and ZED 2]
  • imu/mag: Calibrated magnetometer data [only ZED 2]
  • atm_press: Atmospheric pressure data [only ZED 2]
  • temperature/imu: Temperature of the IMU sensor [only ZED 2]
  • temperature/left: Temperature of the left camera sensor [only ZED 2]
  • temperature/right: Temperature of the right camera sensor [only ZED 2]
  • left_cam_imu_transform: Transform from left camera sensor to IMU sensor position

• Object Detection [only ZED 2]

  • objects: array of the detected/tracked objects for each camera frame [only ZED 2]
  • object_markers: array of markers of the detected/tracked objects to be rendered in Rviz [only ZED 2]

• Diagnostic

  • /diagnostics: ROS diagnostic message for ZED cameras

ZED parameters

You can specify the parameters to be used by the ZED node modifying the values in the files

• params/common.yaml: common paramters to all camera models
• params/zed.yaml: parameters for the ZED camera
• params/zedm.yaml: parameters for the ZED Mini camera
• params/zed2.yaml: parameters for the ZED 2 camera
• params/zed2i.yaml: parameters for the ZED 2i camera

GENERAL PARAMETERS

Parameters with prefix general

PARAMETER	DESCRIPTION	VALUE
camera_name	A custom name for the ZED camera. Used as namespace and prefix for camera TF frames	string, default=zed
camera_model	Type of Stereolabs camera	zed: ZED, zedm: ZED-M, zed2: ZED 2, zed2i: ZED 2i
camera_flip	Flip the camera data if it is mounted upsidedown	true, false
zed_id	Select a ZED camera by its ID. Useful when multiple cameras are connected. ID is ignored if an SVO path is specified	int, default 0
serial_number	Select a ZED camera by its serial number	int, default 0
resolution	Set ZED camera resolution	0: HD2K, 1: HD1080, 2: HD720, 3: VGA
grab_frame_rate	Set ZED camera grabbing framerate	int
gpu_id	Select a GPU device for depth computation	int, default -1 (best device found)
base_frame	Frame_id of the frame that indicates the reference base of the robot	string, default=base_link
verbose	Enable/disable the verbosity of the SDK	true, false
svo_compression	Set SVO compression mode	0: LOSSLESS (PNG/ZSTD), 1: H264 (AVCHD) ,2: H265 (HEVC)
self_calib	Enable/disable self calibration at starting	true, false

VIDEO PARAMETERS

Parameters with prefix video

PARAMETER	DESCRIPTION	VALUE
img_downsample_factor	Resample factor for images [0.01,1.0]. The SDK works with native image sizes, but publishes rescaled image.	double, default=1.0
extrinsic_in_camera_frame	If false extrinsic parameter in camera_info will use ROS native frame (X FORWARD, Z UP) instead of the camera frame (Z FORWARD, Y DOWN) [true use old behavior as for version < v3.1]	true, false

DEPTH PARAMETERS

Parameters with prefix depth

PARAMETER	DESCRIPTION	VALUE
quality	Select depth map quality	0: NONE, 1: PERFORMANCE, 2: MEDIUM, 3: QUALITY, 4: ULTRA
sensing_mode	Select depth sensing mode (change only for VR/AR applications)	0: STANDARD, 1: FILL
depth_stabilization	Enable depth stabilization. Stabilizing the depth requires an additional computation load as it enables tracking	0: disabled, 1: enabled
openni_depth_mode	Convert 32bit depth in meters to 16bit in millimeters	0: 32bit float meters, 1: 16bit uchar millimeters
depth_downsample_factor	Resample factor for depth data matrices [0.01,1.0]. The SDK works with native data sizes, but publishes rescaled matrices (depth map, point cloud, …)	double, default=1.0
min_depth	Minimum value allowed for depth measures	Min: 0.3 (ZED) or 0.1 (ZED-M), Max: 3.0 - Note: reducing this value will require more computational power and GPU memory. In cases of limited compute power, increasing this value can provide better performance
max_depth	Maximum value allowed for depth measures	Min: 1.0, Max: 30.0 - Values beyond this limit will be reported as TOO_FAR

POSITION PARAMETERS

Parameters with prefix pos_tracking

PARAMETER	DESCRIPTION	VALUE
pos_tracking_enabled	Enable/disable positional tracking from start	true, false
publish_tf	Enable/disable publish TF frames	true, false
publish_map_tf	Enable/disable publish map TF frame	true, false
map_frame	Frame_id of the pose message	string, default=map
odometry_frame	Frame_id of the odom message	string, default=odom
area_memory_db_path	Path of the database file for loop closure and relocalization that contains learnt visual information about the environment	string, default=``
save_area_memory_db_on_exit	Save the “known visual features” map when the node is correctly closed to the path indicated by area_memory_db_path	true, false
area_memory	Enable Loop Closing	true, false
floor_alignment	Indicates if the floor must be used as origin for height measures	true, false
initial_base_pose	Initial reference pose	vector, default=[0.0,0.0,0.0, 0.0,0.0,0.0] -> [X, Y, Z, R, P, Y]
init_odom_with_first_valid_pose	Indicates if the odometry must be initialized with the first valid pose received by the tracking algorithm	true, false
path_pub_rate	Frequency (Hz) of publishing of the trajectory messages	float, default=2.0
path_max_count	Maximum number of poses kept in the pose arrays (-1 for infinite)	int, default=-1
two_d_mode	Force navigation on a plane. If true the Z value will be fixed to “fixed_z_value”, roll and pitch to zero	true, false
fixed_z_value:	Value to be used for Z coordinate if two_d_mode is true	double, default: 0.0

MAPPING PARAMETERS

Parameters with prefix mapping

PARAMETER	DESCRIPTION	VALUE
mapping_enabled	Enable/disable the mapping module	true, false
resolution	Resolution of the fused point cloud [0.01, 0.2]	double, default=0.1
max_mapping_range	Maximum depth range while mapping in meters (-1 for automatic calculation) [2.0, 20.0]	double, default=-1
fused_pointcloud_freq	Publishing frequency (Hz) of the 3D map as fused point cloud	double, default=1.0

SENSORS PARAMETERS (ONLY ZED-M, ZED 2, AND ZED 2I)

Parameters with prefix sensors

PARAMETER	DESCRIPTION	VALUE
sensors_timestamp_sync	Synchronize Sensors message timestamp with latest received frame	true, false
publish_imu_tf	Publish IMU -> <cam_name>_left_camera_frame TF	true, false

OBJECT DETECTION PARAMETERS (ONLY ZED 2 AND ZED 2I)

Parameters with prefix object_detection

PARAMETER	DESCRIPTION	VALUE
od_enabled	Enable/disable the Object Detection module when the node starts	true, false
model	Object Detection module to be used	‘0’: MULTI_CLASS_BOX - ‘1’: MULTI_CLASS_BOX_ACCURATE - ‘2’: HUMAN_BODY_FAST - ‘3’: HUMAN_BODY_ACCURATE - ‘4’: MULTI_CLASS_BOX_MEDIUM - ‘5’: HUMAN_BODY_MEDIUM
confidence_threshold	Minimum value of the detection confidence of an object	int [0,100]
max_range:	Maximum detection range	double, default: 15.
object_tracking_enabled	Enable/disable the tracking of the detected objects	true, false
body_fitting:	Enable/disable body fitting for ‘HUMAN_BODY’ models	true, false
mc_people:	Enable/disable the detection of persons for ‘MULTI_CLASS_BOX’ models	true, false
mc_vehicle:	Enable/disable the detection of vehicles for ‘MULTI_CLASS_BOX’ models	true, false
mc_bag:	Enable/disable the detection of bags for ‘MULTI_CLASS_BOX’ models	true, false
mc_animal:	Enable/disable the detection of animals for ‘MULTI_CLASS_BOX’ models	true, false
mc_electronics:	Enable/disable the detection of electronic devices for ‘MULTI_CLASS_BOX’ models	true, false
mc_fruit_vegetable:	Enable/disable the detection of fruits and vegetables for ‘MULTI_CLASS_BOX’ models	true, false

Dynamic parameters

The ZED node lets you reconfigure these parameters dynamically:

PARAMETER	DESCRIPTION	VALUE
pub_frame_rate	Frequency of the publishing of Video and Depth images (equal or minor to grab_frame_rate value)	float [0.1,100.0]
depth_confidence	Threshold to reject depth values based on their confidence. Each depth pixel has a corresponding confidence. A lower value means more confidence and precision (but less density). An upper value reduces filtering (more density, less certainty). A value of 100 will allow values from 0 to 100. (no filtering). A value of 90 will allow values from 10 to 100. (filtering lowest confidence values). A value of 30 will allow values from 70 to 100. (keeping highest confidence values and lowering the density of the depth map). The value should be in [1,100]. By default, the confidence threshold is set at 100, meaning that no depth pixel will be rejected.	int [0,100]
depth_texture_conf	Threshold to reject depth values based on their textureness confidence. A lower value means more confidence and precision (but less density). An upper value reduces filtering (more density, less certainty). The value should be in [1,100]. By default, the confidence threshold is set at 100, meaning that no depth pixel will be rejected.	int [0,100]
point_cloud_freq	Frequency of the pointcloud publishing (equal or minor to frame_rate value)	float [0.1,100.0]
brightness	Defines the brightness control	int [0,8]
contrast	Defines the contrast control	int [0,8]
hue	Defines the hue control	int [0,11]
saturation	Defines the saturation control	int [0,8]
sharpness	Defines the sharpness control	int [0,8]
gamma	Defines the gamma control	int [1,9]
auto_exposure_gain	Defines if the Gain and Exposure are in automatic mode or not	true, false
gain	Defines the gain control [only if auto_exposure_gain is false]	int [0,100]
exposure	Defines the exposure control [only if auto_exposure_gain is false]	int [0,100]
auto_whitebalance	Defines if the White balance is in automatic mode or not	true, false
whitebalance_temperature	Defines the color temperature value (x100)	int [42,65]

To modify a dynamic parameter, you can use the GUI provided by the rqt stack:

$ rosrun rqt_reconfigure rqt_reconfigure

 

Transform frames

The ZED ROS wrapper broadcasts multiple coordinate frames that each provide information about the camera’s position and orientation. If needed, the reference frames can be changed in the launch file.

• base_frame is the current position and orientation of the reference base of the robot
• <camera_name>_camera_center is the current position and orientation of ZED, determined by visual odometry and the tracking module
• <camera_name>_right_camera is the position and orientation of the ZED’s right camera
• <camera_name>_right_camera_optical is the position and orientation of the ZED’s right camera optical frame
• <camera_name>_left_camera is the position and orientation of the ZED’s left camera
• <camera_name>_left_camera_optical is the position and orientation of the ZED’s left camera optical frame
• <camera_name>_imu_link is the origin of the inertial data frame (ZED-M and ZED2 only)
• <camera_name>_mag_link is the origin of the magnetometer frame (ZED2 only)
• <camera_name>_baro_link is the origin of the barometer frame (ZED2 only)
• <camera_name>_temp_left_link is the origin of the left temperature frame (ZED2 only)
• <camera_name>_temp_right_link is the origin of the right temperature frame (ZED2 only)

For RVIZ compatibilty, the root frame map_frame is called map. The TF tree generated by the zed_wrapper reflects the standard descripted in REP105. The odometry frame is updated using only the “visual odometry” information. The map frame is updated using the Tracking algorithm provided by the Stereolabs SDK, fusing the inertial information from the IMU sensor if using a ZED-M camera.

map_frame (`map`)
└─odometry_frame (`odom`)
   └─base_frame (`base_link`)
      └─camera_frame (`<camera_name>_camera_center`)
      |  └─left_camera_frame (`<camera_name>_left_camera_frame`)
      |  |     └─left_camera_optical_frame (`<camera_name>_left_camera_optical_frame`)
      |  |     └─left_temperature:frame (`<camera_name>_temp_left_link`)
      |  └─right_camera_frame (`<camera_name>_right_camera_frame`) (*only ZED 2 and ZED2i*)
      |        └─right_camera_optical_frame (`<camera_name>_right_camera_optical_frame`)
      |        └─left_temperature:frame (`<camera_name>_temp_left_link`)
      └─imu_frame (`<camera_name>_imu_link`) (*only ZED-M and ZED 2 and ZED2i*)
      └─magnetometer_frame (`<camera_name>_mag_link`) (*only ZED 2 and ZED2i*)
      └─barometer_frame (`<camera_name>_baro_link`) (*only ZED 2 and ZED2i*)

ZED-M

The ZED-M provides the same information as the ZED, plus the inertial data from the IMU sensor. The IMU data are used internally to generate the pose in the Map frame with the Tracking sensor fusion algorithm.

Note: The initial pose in Odometry frame can be set to the first pose received by the Tracking module by setting the parameter init_odom_with_first_valid_pose to true.

ZED 2 / ZED 2i

The ZED 2 and ZED 2i provide the same information as the ZED and the ZED-M, plus the data from a new set of sensors. The ZED 2 and ZED 2i provide magnetometer data to get better and absolute information about the YAW angle, atmospheric pressure information to be used to estimate the height of the camera respect to a reference point and temperature information to check the camera health in a complex robotic system.

Services

The ZED node provides the following services:

• reset_tracking: Restarts the Tracking module setting the initial pose to the value available in the param server.
• reset_odometry: Resets the odometry values eliminating the drift due to the Visual Odometry algorithm, setting the new odometry value to the latest camera pose received from the tracking module.
• set_pose: Sets the current pose of the camera to the value passed as single parameters -> X, Y, Z [m], R, P, Y [rad].
• save_area_memory: Triggers the saving of the current area memory used for relocation, if area memory is enabled.
• set_led_status: Sets the status of the blue led -> True: LED ON, False: LED OFF.
• toggle_led: Toggles the status of the blue led, returning the new status.
• start_svo_recording: Starts recording an SVO file. If no filename is provided the default zed.svo is used. If no path is provided with the filename the default recording folder is ~/.ros/.
• stop_svo_recording: Stops an active SVO recording.
• start_remote_stream: Starts streaming over network to allow processing of ZED data on a remote machine. See Remote streaming.
• stop_remote_stream: Stops streaming over network.
• start_3d_mapping: Starts the Spatial Mapping processing. See Spatial Mapping.
• stop_3d_mapping: Stops the Spatial Mapping processing (works even with automatic start from configuration file).
• save_3d_map: Triggers the saving of the current 3D fused point cloud if the mapping module is enabled.
• start_object_detection: Starts the Object Detection processing. See Object Detection Note: returns error if not using a ZED 2 camera.
• stop_object_detection: Stops the Object Detection processing (works even with automatic start from configuration file).

Note: Currently the H26x SVO compression uses the resolution and the real FPS to calculate the bitrate. This feature can lead to some issues of quality when FPS is low. On Nvidia Jetson TX1 and Jetson TX2, the FPS is quite low at the beginning of the grab loop, therefore it could be better to wait for some grab calls before calling start_svo_recording so that the camera FPS is stabilized at the requested value (15fps or more).

Remote streaming

With SDK v2.8 has been introduced the capability to acquire ZED data and to stream it on a remote machine over network. This features is useful when the ZED is connected to a machine with limited CUDA capabilities while a high definition analysis is required. Starting the streaming without activating any other feature (i.e. depth processing, positional tracking, point cloud publishing, …) requires only the power for H264 or H265 compression.

To start streaming the service start_remote_stream must be called with the following parameters:

• codec (def. 0): Defines the codec used for streaming (0: AVCHD [H264], 1: HEVC [H265]) Note: If HEVC (H265) is used, make sure the recieving host is compatible with HEVC decoding (basically a pascal NVIDIA card). If not, prefer to use AVCHD (H264) since every compatible NVIDIA card supports AVCHD decoding
• port (def. 30000): Defines the PORT the data will be streamed on. Note: port must be an even number. Any odd number will be rejected uint16 port=30000
• bitrate (def. 2000): Defines the streaming BITRATE in Kbits/s
• gop_size (def. -1): Defines the GOP SIZE in frame unit. Note: if value is set to -1, the gop size will match 2 seconds, depending on camera fps. The gop size determines the maximum distance between IDR/I-frames. Very high GOP size will result in slightly more efficient compression, especially on static scene. But it can result in more latency if IDR/I-frame packet are lost during streaming. Maximum allowed value is 256 (frames)
• adaptative_bitrate (def. False): Enable/Disable adaptive bitrate. Note: Bitrate will be adjusted regarding the number of packet loss during streaming. If activated, bitrate can vary between [bitrate/4, bitrate]. Currently, the adaptive bitrate only works when “sending” device is a NVIDIA jetson (X1,X2,Xavier,Nano)

RECEIVE A REMOTE STREAM

To acquire the streaming on a remote machine start a ZED node with the following command:

$ roslaunch zed_wrapper zed.launch stream:=<sender_IP>:<port>

For example:

$ roslaunch zed_wrapper zed.launch stream:=192.168.1.127:30000

Note: the stream contains only visual information. Using a ZED-M camera the inertial topic will not be available if not subscribed using ROS standard methods.

Spatial Mapping

The ZED node provides a basilar mapping module publishing a 3D map of the environment as 3D color point cloud (mapping/fused_cloud).

Spatial Mapping is disabled by default as it’s an heavy consuming process, it can be enabled setting to true the parameter mapping/mapping_enabled in the file params/common.yaml.

Spatial Mapping can be enabled manually at any time using the start_3d_mapping service.

To reduce the computational power requirements use value higher than 0.1 m for resolution_m and decrease the value of fused_pointcloud_freq to reduce the frequency of the point cloud topic publishing.

It is possible to save the generated 3D map as a pointcloud in format OBJ, OBJ binary or PLY, by calling the service save_3d_map.

Object Detection

Using the ZED 2 or ZED 2i camera unlocks the Object Detection module that allows to detect and track Human skeletons and Objects.

Object detection is disabled by default as it’s an heavy consuming process, it can be automatically enabled when the node starts setting to true the parameter object_detection/od_enabled in the file params/zed2.yaml or params/zed2i.yaml.

If Object Detection is disabled it can be manually enabled at any time using the start_object_detection service. See Object Detection parameters for more info about the meaning of each parameter used with the service.

The list of the detected objects is published as a custom topic of type ObjectStamped, defined in the package zed_interfaces. For details about the topic fileds please read the Object Detection full documentation.

OBJECT TRACKING

Enabling Object Tracking allows to get information frame by frame about the status of each detected object. The same object is identified with the same ID in all the frames after its first detection and its position is estimated if an occlusion occurs.

Each detected object can get four different values about their Tracking Status:

• OFF - The tracking is not yet initialized, the object ID is not usable
• OK - The object is tracked
• SEARCHING - The object couldn’t be detected in the image and is potentially occluded, the trajectory is estimated
• TERMINATE - This is the last searching state of the track, the track will be deleted in the next retreiveObject

Diagnostic

The ZED node publishes diagnostic information that can be used by the robotics system using a diagnostic_aggregator node.

Using the Runtime monitor plugin of rqt it is possible to get all the diagnostic information and check that the node is working as expected: 

A brief explanation of each field:

• Component: name of the diagnostic component
• Message: summary of the status of the ZED node
• HardwareID: Model of the ZED camera and its serial number
• Capture: grabbing frequency (if video or depth data are subscribed) and the percentage respect to the camera frame rate
• Processing time: time in seconds spent to elaborate data and the time limit to achieve max frame rate
• Playing SVO: (visible only if playng an SVO file) current frame position in the SVO file over the total frame count
• Depth status: indicates if the depth processing is performed
• Point Cloud: point cloud publishing frequency (if there is at least a subscriber) and the percentage respect to the camera frame rate
• Floor Detection: if the floor detection is enabled, indicates if the floor has been detected and the camera position correctly initialized
• Tracking status: indicates the status of the positional tracking, if enabled
• Object data processing: if Object Detection is enabled indicates the time required to process the data relative to detected objects
• IMU: the publishing frequency of the IMU topics, if the camera is the ZED Mini and there is at least a subscriber
• Left CMOS Temp.: (only ZED 2) the temperature of the CMOS of the left camera sensor [-273.15°C if not valid]
• Right CMOS Temp.: (only ZED 2) the temperature of the CMOS of the right camera sensor [-273.15°C if not valid]
• SVO Recording: indicates if the SVO recording is active
• SVO Compression time: average time spent on frame compressing
• SVO Compression ratio: average frame compression ratio

Using multiple ZEDs

It is possible to use multiple ZED cameras with ROS. Simply launch the node with the zed_multi_cam.launch file:

$ roslaunch zed_wrapper zed_multi_cam.launch

Assigning a GPU to a camera

To improve performance, you can specify the gpu_id of the graphic card that will be used for the depth computation in the launch file. The default value (-1) will select the GPU with the highest number of CUDA cores. When using multiple ZEDs, you can assign each camera to a GPU to increase performance.

Limitations

PERFORMANCE

This wrapper lets you quickly prototype applications and interface the ZED with other sensors and packages available in ROS. However, the ROS layer introduces significant latency and a performance hit. If performance is a major concern for your application, please consider using the ZED SDK library.

MULTIPLE ZEDS

The ZED camera uses the maximum bandwidth provided by USB 3.0 to output video. When using multiple ZEDs, you may need to reduce camera framerate and resolution to avoid corrupted frames (green or purple frames). You can also use multiple GPUs to load-balance computations and improve performance.