现有一通过摄像头结合yolo获取图像进行识别定位再通过机械臂运动控制程序进行臂迹规划进行抓取的项目，但是该项目中有一个现象是当机械臂开始运动时摄像头输出画面会自动锁死图像不再实时获取图像，以下是运动控制代码#!/usr/bin/python3
import os
import math
import copy
import json
import actionlib
import control_msgs.msg
from controller import ArmController
from gazebo_msgs.msg import ModelStates
import rospy
from pyquaternion import Quaternion as PyQuaternion
import numpy as np
from gazebo_ros_link_attacher.srv import SetStatic, SetStaticRequest, SetStaticResponse
from gazebo_ros_link_attacher.srv import Attach, AttachRequest, AttachResponse
PKG_PATH = os.path.dirname(os.path.abspath(__file__))
MODELS_INFO = {
"X1-Y2-Z1": {
"home": [0.264589, -0.293903, 0.777]
},
"X2-Y2-Z2": {
"home": [0.277866, -0.724482, 0.777]
},
"X1-Y3-Z2": {
"home": [0.268053, -0.513924, 0.777]
},
"X1-Y2-Z2": {
"home": [0.429198, -0.293903, 0.777]
},
"X1-Y2-Z2-CHAMFER": {
"home": [0.592619, -0.293903, 0.777]
},
"X1-Y4-Z2": {
"home": [0.108812, -0.716057, 0.777]
},
"X1-Y1-Z2": {
"home": [0.088808, -0.295820, 0.777]
},
"X1-Y2-Z2-TWINFILLET": {
"home": [0.103547, -0.501132, 0.777]
},
"X1-Y3-Z2-FILLET": {
"home": [0.433739, -0.507130, 0.777]
},
"X1-Y4-Z1": {
"home": [0.589908, -0.501033, 0.777]
},
"X2-Y2-Z2-FILLET": {
"home": [0.442505, -0.727271, 0.777]
}
}
for model, model_info in MODELS_INFO.items():
pass
#MODELS_INFO[model]["home"] = model_info["home"] + np.array([0.0, 0.10, 0.0])
for model, info in MODELS_INFO.items():
model_json_path = os.path.join(PKG_PATH, "..", "models", f"lego_{model}", "model.json")
# make path absolute
model_json_path = os.path.abspath(model_json_path)
# check path exists
if not os.path.exists(model_json_path):
raise FileNotFoundError(f"Model file {model_json_path} not found")
model_json = json.load(open(model_json_path, "r"))
corners = np.array(model_json["corners"])
size_x = (np.max(corners[:, 0]) - np.min(corners[:, 0]))
size_y = (np.max(corners[:, 1]) - np.min(corners[:, 1]))
size_z = (np.max(corners[:, 2]) - np.min(corners[:, 2]))
#print(f"{model}: {size_x:.3f} x {size_y:.3f} x {size_z:.3f}")
MODELS_INFO[model]["size"] = (size_x, size_y, size_z)
# Compensate for the interlocking height
INTERLOCKING_OFFSET = 0.019
SAFE_X = -0.40
SAFE_Y = -0.13
SURFACE_Z = 0.774
# Resting orientation of the end effector
DEFAULT_QUAT = PyQuaternion(axis=(0, 1, 0), angle=math.pi)
# Resting position of the end effector
DEFAULT_POS = (-0.1, -0.2, 1.2)
DEFAULT_PATH_TOLERANCE = control_msgs.msg.JointTolerance()
DEFAULT_PATH_TOLERANCE.name = "path_tolerance"
DEFAULT_PATH_TOLERANCE.velocity = 10
def get_gazebo_model_name(model_name, vision_model_pose):
"""
Get the name of the model inside gazebo. It is needed for link attacher plugin.
"""
models = rospy.wait_for_message("/gazebo/model_states", ModelStates, timeout=None)
epsilon = 0.05
for gazebo_model_name, model_pose in zip(models.name, models.pose):
if model_name not in gazebo_model_name:
continue
# Get everything inside a square of side epsilon centered in vision_model_pose
ds = abs(model_pose.position.x - vision_model_pose.position.x) + abs(model_pose.position.y - vision_model_pose.position.y)
if ds <= epsilon:
return gazebo_model_name
raise ValueError(f"Model {model_name} at position {vision_model_pose.position.x} {vision_model_pose.position.y} was not found!")
def get_model_name(gazebo_model_name):
return gazebo_model_name.replace("lego_", "").split("_", maxsplit=1)[0]
def get_legos_pos(vision=False):
#get legos position reading vision topic
if vision:
# 从视觉识别主题获取数据
legos = rospy.wait_for_message("/lego_detections", ModelStates, timeout=None)
else:
# 从传感器主题获取数据
models = rospy.wait_for_message("/gazebo/model_states", ModelStates, timeout=None)
legos = ModelStates()
for name, pose in zip(models.name, models.pose):
if "X" not in name:
continue
name = get_model_name(name)
legos.name.append(name)
legos.pose.append(pose)
return [(lego_name, lego_pose) for lego_name, lego_pose in zip(legos.name, legos.pose)]
def straighten(model_pose, gazebo_model_name):
x = model_pose.position.x
y = model_pose.position.y
z = model_pose.position.z
model_quat = PyQuaternion(
x=model_pose.orientation.x,
y=model_pose.orientation.y,
z=model_pose.orientation.z,
w=model_pose.orientation.w)
model_size = MODELS_INFO[get_model_name(gazebo_model_name)]["size"]
"""
Calculate approach quaternion and target quaternion
"""
facing_direction = get_axis_facing_camera(model_quat)
approach_angle = get_approach_angle(model_quat, facing_direction)
print(f"Lego is facing {facing_direction}")
print(f"Angle of approaching measures {approach_angle:.2f} deg")
# Calculate approach quat
approach_quat = get_approach_quat(facing_direction, approach_angle)
# Get above the object
controller.move_to(x, y, target_quat=approach_quat)
# Calculate target quat
regrip_quat = DEFAULT_QUAT
if facing_direction == (1, 0, 0) or facing_direction == (0, 1, 0): # Side
target_quat = DEFAULT_QUAT
pitch_angle = -math.pi/2 + 0.2
if abs(approach_angle) < math.pi/2:
target_quat = target_quat * PyQuaternion(axis=(0, 0, 1), angle=math.pi/2)
else:
target_quat = target_quat * PyQuaternion(axis=(0, 0, 1), angle=-math.pi/2)
target_quat = PyQuaternion(axis=(0, 1, 0), angle=pitch_angle) * target_quat
if facing_direction == (0, 1, 0):
regrip_quat = PyQuaternion(axis=(0, 0, 1), angle=math.pi/2) * regrip_quat
elif facing_direction == (0, 0, -1):
"""
Pre-positioning
"""
controller.move_to(z=z, target_quat=approach_quat)
close_gripper(gazebo_model_name, model_size[0])
tmp_quat = PyQuaternion(axis=(0, 0, 1), angle=2*math.pi/6) * DEFAULT_QUAT
controller.move_to(SAFE_X, SAFE_Y, z+0.05, target_quat=tmp_quat, z_raise=0.1) # Move to safe position
controller.move_to(z=z)
open_gripper(gazebo_model_name)
approach_quat = tmp_quat * PyQuaternion(axis=(1, 0, 0), angle=math.pi/2)
target_quat = approach_quat * PyQuaternion(axis=(0, 0, 1), angle=-math.pi) # Add a yaw rotation of 180 deg
regrip_quat = tmp_quat * PyQuaternion(axis=(0, 0, 1), angle=math.pi)
else:
target_quat = DEFAULT_QUAT
target_quat = target_quat * PyQuaternion(axis=(0, 0, 1), angle=-math.pi/2)
"""
Grip the model
"""
if facing_direction == (0, 0, 1) or facing_direction == (0, 0, -1):
closure = model_size[0]
z = SURFACE_Z + model_size[2] / 2
elif facing_direction == (1, 0, 0):
closure = model_size[1]
z = SURFACE_Z + model_size[0] / 2
elif facing_direction == (0, 1, 0):
closure = model_size[0]
z = SURFACE_Z + model_size[1] / 2
controller.move_to(z=z, target_quat=approach_quat)
close_gripper(gazebo_model_name, closure)
"""
Straighten model if needed
"""
if facing_direction != (0, 0, 1):
z = SURFACE_Z + model_size[2]/2
controller.move_to(z=z+0.05, target_quat=target_quat, z_raise=0.1)
controller.move(dz=-0.05)
open_gripper(gazebo_model_name)
# Re grip the model
controller.move_to(z=z, target_quat=regrip_quat, z_raise=0.1)
close_gripper(gazebo_model_name, model_size[0])
def close_gripper(gazebo_model_name, closure=0):
set_gripper(0.81-closure*10)
rospy.sleep(0.5)
# Create dynamic joint
if gazebo_model_name is not None:
req = AttachRequest()
req.model_name_1 = gazebo_model_name
req.link_name_1 = "link"
req.model_name_2 = "robot"
req.link_name_2 = "wrist_3_link"
attach_srv.call(req)
def open_gripper(gazebo_model_name=None):
set_gripper(0.0)
# Destroy dynamic joint
if gazebo_model_name is not None:
req = AttachRequest()
req.model_name_1 = gazebo_model_name
req.link_name_1 = "link"
req.model_name_2 = "robot"
req.link_name_2 = "wrist_3_link"
detach_srv.call(req)
def set_model_fixed(model_name):
req = AttachRequest()
req.model_name_1 = model_name
req.link_name_1 = "link"
req.model_name_2 = "ground_plane"
req.link_name_2 = "link"
attach_srv.call(req)
req = SetStaticRequest()
print("{} TO HOME".format(model_name))
req.model_name = model_name
req.link_name = "link"
req.set_static = True
setstatic_srv.call(req)
def get_approach_quat(facing_direction, approach_angle):
quat = DEFAULT_QUAT
if facing_direction == (0, 0, 1):
pitch_angle = 0
yaw_angle = 0
elif facing_direction == (1, 0, 0) or facing_direction == (0, 1, 0):
pitch_angle = + 0.2
if abs(approach_angle) < math.pi/2:
yaw_angle = math.pi/2
else:
yaw_angle = -math.pi/2
elif facing_direction == (0, 0, -1):
pitch_angle = 0
yaw_angle = 0
else:
raise ValueError(f"Invalid model state {facing_direction}")
quat = quat * PyQuaternion(axis=(0, 1, 0), angle=pitch_angle)
quat = quat * PyQuaternion(axis=(0, 0, 1), angle=yaw_angle)
quat = PyQuaternion(axis=(0, 0, 1), angle=approach_angle+math.pi/2) * quat
return quat
def get_axis_facing_camera(quat):
axis_x = np.array([1, 0, 0])
axis_y = np.array([0, 1, 0])
axis_z = np.array([0, 0, 1])
new_axis_x = quat.rotate(axis_x)
new_axis_y = quat.rotate(axis_y)
new_axis_z = quat.rotate(axis_z)
# get angle between new_axis and axis_z
angle = np.arccos(np.clip(np.dot(new_axis_z, axis_z), -1.0, 1.0))
# get if model is facing up, down or sideways
if angle < np.pi / 3:
return 0, 0, 1
elif angle < np.pi / 3 * 2 * 1.2:
if abs(new_axis_x[2]) > abs(new_axis_y[2]):
return 1, 0, 0
else:
return 0, 1, 0
#else:
# raise Exception(f"Invalid axis {new_axis_x}")
else:
return 0, 0, -1
def get_approach_angle(model_quat, facing_direction):#get gripper approach angle
if facing_direction == (0, 0, 1):
return model_quat.yaw_pitch_roll[0] - math.pi/2 #rotate gripper
elif facing_direction == (1, 0, 0) or facing_direction == (0, 1, 0):
axis_x = np.array([0, 1, 0])
axis_y = np.array([-1, 0, 0])
new_axis_z = model_quat.rotate(np.array([0, 0, 1])) #get z axis of lego
# get angle between new_axis and axis_x
dot = np.clip(np.dot(new_axis_z, axis_x), -1.0, 1.0) #sin angle between lego z axis and x axis in fixed frame
det = np.clip(np.dot(new_axis_z, axis_y), -1.0, 1.0) #cos angle between lego z axis and x axis in fixed frame
return math.atan2(det, dot) #get angle between lego z axis and x axis in fixed frame
elif facing_direction == (0, 0, -1):
return -(model_quat.yaw_pitch_roll[0] - math.pi/2) % math.pi - math.pi
else:
raise ValueError(f"Invalid model state {facing_direction}")
def set_gripper(value):
goal = control_msgs.msg.GripperCommandGoal()
goal.command.position = value # From 0.0 to 0.8
goal.command.max_effort = -1 # # Do not limit the effort
action_gripper.send_goal_and_wait(goal, rospy.Duration(10))
return action_gripper.get_result()
if __name__ == "__main__":
print("Initializing node of kinematics")
rospy.init_node("send_joints")
controller = ArmController()
# Create an action client for the gripper
action_gripper = actionlib.SimpleActionClient(
"/gripper_controller/gripper_cmd",
control_msgs.msg.GripperCommandAction
)
print("Waiting for action of gripper controller")
action_gripper.wait_for_server()
setstatic_srv = rospy.ServiceProxy("/link_attacher_node/setstatic", SetStatic)
attach_srv = rospy.ServiceProxy("/link_attacher_node/attach", Attach)
detach_srv = rospy.ServiceProxy("/link_attacher_node/detach", Attach)
setstatic_srv.wait_for_service()
attach_srv.wait_for_service()
detach_srv.wait_for_service()
controller.move_to(*DEFAULT_POS, DEFAULT_QUAT)
print("Waiting for detection of the models")
rospy.sleep(0.5)
legos = get_legos_pos(vision=True)
legos.sort(reverse=True, key=lambda a: (a[1].position.x, a[1].position.y))
for model_name, model_pose in legos:
open_gripper()
try:
model_home = MODELS_INFO[model_name]["home"]
model_size = MODELS_INFO[model_name]["size"]
except ValueError as e:
print(f"Model name {model_name} was not recognized!")
continue
# Get actual model_name at model xyz coordinates
try:
gazebo_model_name = get_gazebo_model_name(model_name, model_pose)
except ValueError as e:
print(e)
continue
# Straighten lego
straighten(model_pose, gazebo_model_name)
controller.move(dz=0.15)
"""
Go to destination
"""
x, y, z = model_home
z += model_size[2] / 2 +0.004
print(f"Moving model {model_name} to {x} {y} {z}")
controller.move_to(x, y, target_quat=DEFAULT_QUAT * PyQuaternion(axis=[0, 0, 1], angle=math.pi / 2))
# Lower the object and release
controller.move_to(x, y, z)
set_model_fixed(gazebo_model_name)
open_gripper(gazebo_model_name)
controller.move(dz=0.15)
if controller.gripper_pose[0][1] > -0.3 and controller.gripper_pose[0][0] > 0:
controller.move_to(*DEFAULT_POS, DEFAULT_QUAT)
# increment z in order to stack lego correctly
MODELS_INFO[model_name]["home"][2] += model_size[2] - INTERLOCKING_OFFSET
print("Moving to Default Position")
controller.move_to(*DEFAULT_POS, DEFAULT_QUAT)
open_gripper()
rospy.sleep(0.4)
以下是视觉识别代码：
#! /usr/bin/env python3
import cv2 as cv
import numpy as np
import torch
import message_filters
import rospy
import sys
import time
from sensor_msgs.msg import Image
from cv_bridge import CvBridge
from rospkg import RosPack # get abs path
from os import path # get home path
from gazebo_msgs.msg import ModelStates
from geometry_msgs.msg import *
from pyquaternion import Quaternion as PyQuaternion
bridge = CvBridge()
# Global variables
path_yolo = path.join(path.expanduser('~'), 'yolov5')
path_vision = RosPack().get_path('vision')
path_weigths = path.join(path_vision, 'weigths')
cam_point = (-0.44, -0.5, 1.58)
height_tavolo = 0.74
dist_tavolo = None
origin = None
model = None
model_orientation = None
legoClasses = ['X1-Y1-Z2', 'X1-Y2-Z1', 'X1-Y2-Z2', 'X1-Y2-Z2-CHAMFER', 'X1-Y2-Z2-TWINFILLET', 'X1-Y3-Z2', 'X1-Y3-Z2-FILLET', 'X1-Y4-Z1', 'X1-Y4-Z2', 'X2-Y2-Z2', 'X2-Y2-Z2-FILLET']
argv = sys.argv
a_show = '-show' in argv
# Utility Functions
def get_dist_tavolo(depth, hsv, img_draw):
global dist_tavolo
#color = (120,1,190)
#mask = get_lego_mask(color, hsv, (5, 5, 5))
#dist_tavolo = depth[mask].max()
#if dist_tavolo > 1: dist_tavolo -= height_tavolo
dist_tavolo = np.nanmax(depth)
def get_origin(img):
global origin
origin = np.array(img.shape[1::-1]) // 2
def get_lego_distance(depth):
return depth.min()
def get_lego_color(center, rgb):
return rgb[center].tolist()
def get_lego_mask(color, hsv, toll = (20, 20, 255)):
thresh = np.array(color)
mintoll = thresh - np.array([toll[0], toll[1], min(thresh[2]-1, toll[2])])
maxtoll = thresh + np.array(toll)
return cv.inRange(hsv, mintoll, maxtoll)
def getDepthAxis(height, lego):
X, Y, Z = (int(x) for x in lego[1:8:3])
#Z = (0.038, 0.057) X = (0.031, 0.063) Y = (0.031, 0.063, 0.095, 0.127)
rapZ = height / 0.019 - 1
pinZ = round(rapZ)
rapXY = height / 0.032
pinXY = round(rapXY)
errZ = abs(pinZ - rapZ) + max(pinZ - 2, 0)
errXY = abs(pinXY - rapXY) + max(pinXY - 4, 0)
if errZ < errXY:
return pinZ, 2, pinZ == Z # pin, is ax Z, match
else:
if pinXY == Y: return pinXY, 1, True
else: return pinXY, 0, pinXY == X
def point_distorption(point, height, origin):
p = dist_tavolo / (dist_tavolo - height)
point = point - origin
return p * point + origin
def point_inverse_distortption(point, height):
p = dist_tavolo / (dist_tavolo - height)
point = point - origin
return point / p + origin
def myimshow(title, img):
def mouseCB(event,x,y,a,b):
print(x, y, img[y, x], "\r",end='',flush=True)
print("\033[K", end='')
cv.imshow(title, img)
cv.setMouseCallback(title, mouseCB)
cv.waitKey()
# ----------------- LOCALIZATION ----------------- #
def process_item(imgs, item):
#images
rgb, hsv, depth, img_draw = imgs
#obtaining Yolo informations (class, coordinates, center)
x1, y1, x2, y2, cn, cl, nm = item.values()
mar = 15
x1, y1 = max(mar, x1), max(mar, y1)
x2, y2 = min(rgb.shape[1]-mar, x2), min(rgb.shape[0]-mar, y2)
boxMin = np.array((x1-mar, y1-mar))
x1, y1, x2, y2 = np.int0((x1, y1, x2, y2))
boxCenter = (y2 + y1) // 2, (x2 + x1) // 2
color = get_lego_color(boxCenter, rgb)
hsvcolor = get_lego_color(boxCenter, hsv)
sliceBox = slice(y1-mar, y2+mar), slice(x1-mar, x2+mar)
#crop img with coordinate bounding box; computing all imgs
l_rgb = rgb[sliceBox]
l_hsv = hsv[sliceBox]
if a_show: cv.rectangle(img_draw, (x1,y1),(x2,y2), color, 2)
l_depth = depth[sliceBox]
l_mask = get_lego_mask(hsvcolor, l_hsv) # filter mask by color
l_mask = np.where(l_depth < dist_tavolo, l_mask, 0)
l_depth = np.where(l_mask != 0, l_depth, dist_tavolo)
#myimshow("asda", hsv)
#getting lego height from camera and table
l_dist = get_lego_distance(l_depth)
l_height = dist_tavolo - l_dist
#masking
l_top_mask = cv.inRange(l_depth, l_dist-0.002, l_dist+0.002)
#cv.bitwise_xor(img_draw,img_draw,img_draw, mask=cv.inRange(depth, l_dist-0.002, l_dist+0.002))
#myimshow("hmask", l_top_mask)
# model detect orientation
depth_borded = np.zeros(depth.shape, dtype=np.float32)
depth_borded[sliceBox] = l_depth
depth_image = cv.normalize(
depth_borded, None, alpha=0, beta=255, norm_type=cv.NORM_MINMAX, dtype=cv.CV_8U
)
depth_image = cv.cvtColor(depth_image, cv.COLOR_GRAY2RGB).astype(np.uint8)
#yolo in order to keep te orientation
#print("Model orientation: Start...", end='\r')
model_orientation.conf = 0.7
results = model_orientation(depth_image)
pandino = []
pandino = results.pandas().xyxy[0].to_dict(orient="records")
n = len(pandino)
#print("Model orientation: Finish", n)
# Adjust prediction
pinN, ax, isCorrect = getDepthAxis(l_height, nm)
if not isCorrect and ax == 2:
if cl in (1,2,3) and pinN in (1, 2): # X1-Y2-Z*, *1,2,2-CHAMFER
cl = 1 if pinN == 1 else 2 # -> Z'pinN'
elif cl in (7, 8) and pinN in (1, 2): # X1-Y4-Z*, *1,2
cl = 7 if pinN == 1 else 8 # -> Z'pinN'
elif pinN == -1:
nm = "{} -> {}".format(nm, "Target")
else:
print("[Warning] Error in classification")
elif not isCorrect:
ax = 1
if cl in (0, 2, 5, 8) and pinN <= 4: # X1-Y*-Z2, *1,2,3,4
cl = (2, 5, 8)[pinN-2] # -> Y'pinN'
elif cl in (1, 7) and pinN in (2, 4): # X1-Y*-Z1, *2,4
cl = 1 if pinN == 2 else 7 # -> Y'pinN'
elif cl == 9 and pinN == 1: # X2-Y2-Z2
cl = 2 # -> X1
ax = 0
else: print("[Warning] Error in classification")
nm = legoClasses[cl]
if n != 1:
print("[Warning] Classification not found")
or_cn, or_cl, or_nm = ['?']*3
or_nm = ('lato', 'lato', 'sopra/sotto')[ax]
else:
print()
#obtaining Yolo informations (class, coordinates, center)
or_item = pandino[0]
or_cn, or_cl, or_nm = or_item['confidence'], or_item['class'], or_item['name']
if or_nm == 'sotto': ax = 2
if or_nm == 'lato' and ax == 2: ax = 1
#---
#creating silouette top surface lego
contours, hierarchy = cv.findContours(l_top_mask, cv.RETR_TREE, cv.CHAIN_APPROX_SIMPLE)
top_center = np.zeros(2)
for cnt in contours:
tmp_center, top_size, top_angle = cv.minAreaRect(cnt)
top_center += np.array(tmp_center)
top_center = boxMin + top_center / len(contours)
#creating silouette lego
contours, hierarchy = cv.findContours(l_mask, cv.RETR_TREE, cv.CHAIN_APPROX_SIMPLE)
if len(contours) == 0: return None
cnt = contours[0]
l_center, l_size, l_angle = cv.minAreaRect(cnt)
l_center += boxMin
if ax != 2: l_angle = top_angle
l_box = cv.boxPoints((l_center, l_size, l_angle))
if l_size[0] <=3 or l_size[1] <=3:
cv.drawContours(img_draw, np.int0([l_box]), 0, (0,0,0), 2)
return None # filter out artifacts
if a_show: cv.drawContours(img_draw, np.int0([l_box]), 0, color, 2)
# silouette distorption
# get vertexs distance from origin
top_box = l_box.copy()
vertexs_norm = [(i, np.linalg.norm(vec - origin)) for vec, i in zip(l_box, range(4))]
vertexs_norm.sort(key=lambda tup: tup[1])
# get closest vertex
iver = vertexs_norm[0][0]
vec = l_box[iver]
# distorping closest vertex
if or_nm == 'sopra': l_height -= 0.019
top_box[iver] = point_distorption(l_box[iver], l_height, origin)
v0 = top_box[iver] - vec
# adapt adiacents veretx
v1 = l_box[iver - 3] - vec # i - 3 = i+1 % 4
v2 = l_box[iver - 1] - vec
top_box[iver - 3] += np.dot(v0, v2) / np.dot(v2, v2) * v2
top_box[iver - 1] += np.dot(v0, v1) / np.dot(v1, v1) * v1
l_center = (top_box[0] + top_box[2]) / 2
if a_show:
cv.drawContours(img_draw, np.int0([top_box]), 0, (5,5,5), 2)
cv.circle(img_draw, np.int0(top_box[iver]),1, (0,0,255),1,cv.LINE_AA)
#rotation and axis drawing
if or_nm in ('sopra', 'sotto', 'sopra/sotto', '?'): # fin x, y directions (dirX, dirY)
dirZ = np.array((0,0,1))
if or_nm == 'sotto': dirZ = np.array((0,0,-1))
projdir = l_center - top_center
if np.linalg.norm(projdir) < l_size[0] / 10:
dirY = top_box[0] - top_box[1]
dirX = top_box[0] - top_box[-1]
if np.linalg.norm(dirY) < np.linalg.norm(dirX): dirX, dirY = dirY, dirX
projdir = dirY * np.dot(dirY, projdir)
edgeFar = [ver for ver in top_box if np.dot(ver - l_center, projdir) >= 0][:2]
dirY = (edgeFar[0] + edgeFar[1]) / 2 - l_center
dirY /= np.linalg.norm(dirY)
dirY = np.array((*dirY, 0))
dirX = np.cross(dirZ, dirY)
elif or_nm == "lato": # find pin direction (dirZ)
edgePin = [ver for ver in top_box if np.dot(ver - l_center, l_center - top_center) >= 0][:2]
dirZ = (edgePin[0] + edgePin[1]) / 2 - l_center
dirZ /= np.linalg.norm(dirZ)
dirZ = np.array((*dirZ, 0))
if cl == 10:
if top_size[1] > top_size[0]: top_size = top_size[::-1]
if top_size[0] / top_size[1] < 1.7: ax = 0
if ax == 0:
vx,vy,x,y = cv.fitLine(cnt, cv.DIST_L2,0,0.01,0.01)
dir = np.array((vx, vy))
vertexs_distance = [abs(np.dot(ver - l_center, dir)) for ver in edgePin]
iverFar = np.array(vertexs_distance).argmin()
dirY = edgePin[iverFar] - edgePin[iverFar-1]
dirY /= np.linalg.norm(dirY)
dirY = np.array((*dirY, 0))
dirX = np.cross(dirZ, dirY)
if a_show: cv.circle(img_draw, np.int0(edgePin[iverFar]), 5, (70,10,50), 1)
#cv.line(img_draw, np.int0(l_center), np.int0(l_center+np.array([int(vx*100),int(vy*100)])),(0,0,255), 3)
if ax == 1:
dirY = np.array((0,0,1))
dirX = np.cross(dirZ, dirY)
if a_show: cv.line(img_draw, *np.int0(edgePin), (255,255,0), 2)
l_center = point_inverse_distortption(l_center, l_height)
# post rotation extra
theta = 0
if cl == 1 and ax == 1: theta = 1.715224 - np.pi / 2
if cl == 3 and or_nm == 'sotto': theta = 2.359515 - np.pi
if cl == 4 and ax == 1: theta = 2.145295 - np.pi
if cl == 6 and or_nm == 'sotto': theta = 2.645291 - np.pi
if cl == 10 and or_nm == 'sotto': theta = 2.496793 - np.pi
rotX = PyQuaternion(axis=dirX, angle=theta)
dirY = rotX.rotate(dirY)
dirZ = rotX.rotate(dirZ)
if a_show:
# draw frame
lenFrame = 50
unit_z = 0.031
unit_x = 22 * 0.8039 / dist_tavolo
x_to_z = lenFrame * unit_z/unit_x
center = np.int0(l_center)
origin_from_top = origin - l_center
endX = point_distorption(lenFrame * dirX[:2], x_to_z * dirX[2], origin_from_top)
frameX = (center, center + np.int0(endX))
endY = point_distorption(lenFrame * dirY[:2], x_to_z * dirY[2], origin_from_top)
frameY = (center, center + np.int0(endY))
endZ = point_distorption(lenFrame * dirZ[:2], x_to_z * dirZ[2], origin_from_top)
frameZ = (center, center + np.int0(endZ))
cv.line(img_draw, *frameX, (0,0,255), 2)
cv.line(img_draw, *frameY, (0,255,0), 2)
cv.line(img_draw, *frameZ, (255,0,0), 2)
# ---
# draw text
if or_cl != '?': or_cn = ['SIDE', 'UP', 'DOWN'][or_cl]
text = "{} {:.2f} {}".format(nm, cn, or_cn)
(text_width, text_height) = cv.getTextSize(text, cv.FONT_HERSHEY_DUPLEX, 0.4, 1)[0]
text_offset_x = boxCenter[1] - text_width // 2
text_offset_y = y1 - text_height
box_coords = ((text_offset_x - 1, text_offset_y + 1), (text_offset_x + text_width + 1, text_offset_y - text_height - 1))
cv.rectangle(img_draw, box_coords[0], box_coords[1], (210,210,10), cv.FILLED)
cv.putText(img_draw, text, (text_offset_x, text_offset_y), cv.FONT_HERSHEY_DUPLEX, 0.4, (255, 255, 255), 1)
def getAngle(vec, ax):
vec = np.array(vec)
if not vec.any(): return 0
vec = vec / np.linalg.norm(vec)
wise = 1 if vec[-1] >= 0 else -1
dotclamp = max(-1, min(1, np.dot(vec, np.array(ax))))
return wise * np.arccos(dotclamp)
msg = ModelStates()
msg.name = nm
#fov = 1.047198
#rap = np.tan(fov)
#print("rap: ", rap)
xyz = np.array((l_center[0], l_center[1], l_height / 2 + height_tavolo))
xyz[:2] /= rgb.shape[1], rgb.shape[0]
xyz[:2] -= 0.5
xyz[:2] *= (-0.968, 0.691)
xyz[:2] *= dist_tavolo / 0.84
xyz[:2] += cam_point[:2]
rdirX, rdirY, rdirZ = dirX, dirY, dirZ
rdirX[0] *= -1
rdirY[0] *= -1
rdirZ[0] *= -1
qz1 = PyQuaternion(axis=(0,0,1), angle=-getAngle(dirZ[:2], (1,0)))
rdirZ = qz1.rotate(dirZ)
qy2 = PyQuaternion(axis=(0,1,0), angle=-getAngle((rdirZ[2],rdirZ[0]), (1,0)))
rdirX = qy2.rotate(qz1.rotate(rdirX))
qz3 = PyQuaternion(axis=(0,0,1), angle=-getAngle(rdirX[:2], (1,0)))
rot = qz3 * qy2 * qz1
rot = rot.inverse
msg.pose = Pose(Point(*xyz), Quaternion(x=rot.x,y=rot.y,z=rot.z,w=rot.w))
#pub.publish(msg)
#print(msg)
return msg
#image processing
def process_image(rgb, depth):
# img_draw = rgb.copy() # 创建RGB图像副本用于可视化[^4]
# hsv = cv.cvtColor(rgb, cv.COLOR_BGR2HSV) # 转换为HSV色彩空间[^4]
# get_dist_tavolo(depth, hsv, img_draw) # 深度信息与HSV结合处理
# get_origin(rgb) # 基于原始RGB图像确定坐标系原点
# #results collecting localization
# #print("Model localization: Start...",end='\r')
# model.conf = 0.6 # 设置检测置信度阈值
# results = model(rgb) # YOLO模型推理[^2]
# pandino = results.pandas().xyxy[0].to_dict(orient="records") # 结果转为字典
# #print("Model localization: Finish ")
# # ----
# if depth is not None:
# imgs = (rgb, hsv, depth, img_draw) # 多源数据组合
# results = [process_item(imgs, item) for item in pandino] # 逐项处理
# # ----
# msg = ModelStates() # 创建ROS消息对象
# for point in results:
# if point is not None:
# msg.name.append(point.name) # 添加物体名称
# msg.pose.append(point.pose) # 添加位姿信息
# pub.publish(msg) #发布信息
# if a_show: # 显示控制标志
# cv.imshow("vision-results" \
# "", img_draw) # 显示结果图像
# cv.waitKey() #等待按键
# 移除原代码中的图像读取和显示等待逻辑
img_draw = rgb.copy()
hsv = cv.cvtColor(rgb, cv.COLOR_BGR2HSV)
# 简化深度处理（根据实际需求调整）
if depth is not None:
get_dist_tavolo(depth, hsv, img_draw)
get_origin(rgb)
# YOLO实时检测（调整置信度平衡速度精度）
model.conf = 0.2 # 可降低至0.4提升速度
results = model(rgb)
# 结果处理
pandino = results.pandas().xyxy[0].to_dict(orient="records")
if depth is not None:
imgs = (rgb, hsv, depth, img_draw)
results = [process_item(imgs, item) for item in pandino]
# 发布结果
msg = ModelStates()
for point in results:
if point is not None:
msg.name.append(point.name)
msg.pose.append(point.pose)
pub.publish(msg)
# 实时显示（可选）
if a_show:
cv.imshow("Camera", img_draw)
cv.waitKey(25) & 0xFF # 关键修改：非阻塞等待
pass
def process_CB(image_rgb, image_depth):
t_start = time.time()
#from standard message image to opencv image
rgb = CvBridge().imgmsg_to_cv2(image_rgb, "bgr8")
depth = CvBridge().imgmsg_to_cv2(image_depth, "32FC1")
process_image(rgb, depth)
print("Time:", time.time() - t_start)
# rospy.signal_shutdown(0)
pass
#init node function
def start_node():
global pub
print("Starting Node Vision 1.0")
rospy.init_node('vision')
print("Subscribing to camera images")
# viewer = ImageViewer()
#topics subscription
rgb = message_filters.Subscriber("/camera/color/image_raw", Image,queue_size=1)
depth = message_filters.Subscriber("/camera/depth/image_raw", Image,queue_size=1)
#publisher results
pub=rospy.Publisher("lego_detections", ModelStates, queue_size=1)
print("Localization is starting.. ")
print("(Waiting for images..)", end='\r'), print(end='\033[K')
#images synchronization
syncro = message_filters.TimeSynchronizer([rgb, depth], 1, reset=True)
syncro.registerCallback(process_CB)
#keep node always alive
rospy.spin()
# Destroy OpenCV windows on shutdown
cv.destroyAllWindows()
pass
def load_models():
global model, model_orientation
#yolo model and weights classification
print("Loading model best.pt")
weight = path.join(path_weigths, 'best.pt')
model = torch.hub.load(path_yolo,'custom',path=weight, source='local')
#yolo model and weights orientation
print("Loading model orientation.pt")
weight = path.join(path_weigths, 'depth.pt')
model_orientation = torch.hub.load(path_yolo,'custom',path=weight, source='local')
pass
# class ImageViewer:
# def __init__(self):
# # Initialize the CvBridge
# self.bridge = CvBridge()
# # Get parameters
# self.view_image = rospy.get_param('~view_image', True)
# self.input_image_topic = rospy.get_param('~input_image_topic', '/camera/color/image_raw')
# # Subscribe to the image topic
# self.image_sub = rospy.Subscriber(self.input_image_topic, Image, self.image_callback)
# def image_callback(self, data):
# if self.view_image:
# # Convert the ROS Image message to OpenCV format
# cv_image = self.bridge.imgmsg_to_cv2(data, "bgr8")
# # Display the image using OpenCV
# cv.imshow("Camera View", cv_image)
# cv.waitKey(1)
def image_callback(msg):
"""ROS图像消息回调函数"""
try:
# 将ROS图像消息转为OpenCV格式(BGR)
cv_image = bridge.imgmsg_to_cv2(msg, "bgr8")
# 调用处理函数
process_image(cv_image, None) # depth设为None
except Exception as e:
rospy.logerr(f"图像转换错误: {e}")
if __name__ == '__main__':
load_models()
try:
start_node()
except rospy.ROSInterruptException:
pass
要如何进行修改
最新发布
本文介绍了一次在IDES中对PO/合同审批策略的修改过程，包括新建审批策略类FRG_EKKO_BOYD，增加了凭证类别和公司代码字段以区分不同类型的审批需求。文章还提到了更新审批策略类的具体步骤及遇到的问题。
扫一扫
1323
被折叠的 条评论
为什么被折叠?
到【灌水乐园】发言
