StarRailCopilot/tasks/map/interact/aim.py
2023-11-07 13:36:45 +08:00

347 lines
9.8 KiB
Python

import cv2
import numpy as np
from module.base.decorator import cached_property, del_cached_property
from module.base.utils import Points, image_size, load_image
from module.config.utils import dict_to_kv
from module.logger import logger
from tasks.base.ui import UI
def inrange(image, lower=0, upper=255):
"""
Get the coordinates of pixels in range.
Equivalent to `np.array(np.where(lower <= image <= upper))` but faster.
Note that this method will change `image`.
`cv2.findNonZero()` is faster than `np.where`
points = np.array(np.where(y > 24)).T[:, ::-1]
points = np.array(cv2.findNonZero((y > 24).astype(np.uint8)))[:, 0, :]
`cv2.inRange(y, 24)` is faster than `y > 24`
cv2.inRange(y, 24, 255, dst=y)
y = y > 24
Returns:
np.ndarray: Shape (N, 2)
E.g. [[x1, y1], [x2, y2], ...]
"""
cv2.inRange(image, lower, upper, dst=image)
try:
return np.array(cv2.findNonZero(image))[:, 0, :]
except IndexError:
# Empty result
# IndexError: too many indices for array: array is 0-dimensional, but 3 were indexed
return np.array([])
def subtract_blur(image, radius=3, negative=False):
"""
If you care performance more than quality:
- radius=3, use medianBlur
- radius=5,7,9,11, use GaussianBlur
- radius>11, use stackBlur (requires opencv >= 4.7.0)
Args:
image:
radius:
negative:
Returns:
np.ndarray:
"""
if radius <= 3:
blur = cv2.medianBlur(image, radius)
elif radius <= 11:
blur = cv2.GaussianBlur(image, (radius, radius), 0)
else:
blur = cv2.stackBlur(image, (radius, radius), 0)
if negative:
cv2.subtract(blur, image, dst=blur)
else:
cv2.subtract(image, blur, dst=blur)
return blur
def remove_border(image, radius):
"""
Paint edge pixels black.
No returns, changes are written to `image`
Args:
image:
radius:
"""
width, height = image_size(image)
image[:, :radius + 1] = 0
image[:, width - radius:] = 0
image[:radius + 1, :] = 0
image[height - radius:, :] = 0
def create_circle(min_radius, max_radius):
"""
Create a circle with min_radius <= R <= max_radius.
1 represents circle, 0 represents background
Args:
min_radius:
max_radius:
Returns:
np.ndarray:
"""
circle = np.ones((max_radius * 2 + 1, max_radius * 2 + 1), dtype=np.uint8)
center = np.array((max_radius, max_radius))
points = np.array(np.meshgrid(np.arange(circle.shape[0]), np.arange(circle.shape[1]))).T
distance = np.linalg.norm(points - center, axis=2)
circle[distance < min_radius] = 0
circle[distance > max_radius] = 0
return circle
def draw_circle(image, circle, points):
"""
Add a circle onto image.
No returns, changes are written to `image`
Args:
image:
circle: Created from create_circle()
points: (x, y), center of the circle to draw
"""
width, height = image_size(circle)
x1 = -int(width // 2)
y1 = -int(height // 2)
x2 = width + x1
y2 = height + y1
for point in points:
x, y = point
# Fancy index is faster
index = image[y + y1:y + y2, x + x1:x + x2]
# print(index.shape)
cv2.add(index, circle, dst=index)
class Aim:
radius_enemy = (24, 25)
radius_item = (8, 10)
def __init__(self):
self.debug = False
self.draw_item = None
self.draw_enemy = None
self.points_item = None
self.points_enemy = None
def clear_image_cache(self):
self.draw_item = None
self.draw_enemy = None
self.points_item = None
self.points_enemy = None
del_cached_property(self, 'aimed_enemy')
del_cached_property(self, 'aimed_item')
@cached_property
def mask_interact(self):
return load_image('./assets/mask/MASK_MAP_INTERACT.png')
@cached_property
def circle_enemy(self):
return create_circle(*self.radius_enemy)
@cached_property
def circle_item(self):
return create_circle(*self.radius_item)
# @timer
def predict_enemy(self, h, v):
min_radius, max_radius = self.radius_enemy
width, height = image_size(v)
# Get white circle `y`
y = subtract_blur(h, 3, negative=False)
cv2.inRange(h, 168, 255, h)
cv2.bitwise_and(y, h, dst=y)
# Get red glow `v`
cv2.inRange(v, 168, 255, dst=v)
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
cv2.dilate(v, kernel, dst=v)
# Remove noise and leave red circle only
cv2.bitwise_and(y, v, dst=y)
# cv2.imshow('predict_enemy', y)
# Remove game UI
cv2.bitwise_and(y, self.mask_interact, dst=y)
# Remove points on the edge, or draw_circle() will overflow
remove_border(y, max_radius)
# Get all pixels
points = inrange(y, lower=18)
if points.shape[0] > 1000:
logger.warning(f'AimDetector.predict_enemy() too many points to draw: {points.shape}')
# Draw circles
draw = np.zeros((height, width), dtype=np.uint8)
draw_circle(draw, self.circle_enemy, points)
if self.debug:
self.draw_enemy = cv2.multiply(draw, 4)
draw = subtract_blur(draw, 3)
# Find peaks
points = inrange(draw, lower=36)
points = Points(points).group(threshold=10)
if points.shape[0] > 3:
logger.warning(f'AimDetector.predict_enemy() too many peaks: {points.shape}')
self.points_enemy = points
# print(points)
return points
# @timer
def predict_item(self, v):
min_radius, max_radius = self.radius_item
width, height = image_size(v)
# Get white circle `y`
y = subtract_blur(v, 9)
white = cv2.inRange(v, 112, 144)
cv2.bitwise_and(y, white, dst=y)
# Get cyan glow `v`
cv2.inRange(v, 0, 84, dst=v)
kernel = cv2.getStructuringElement(cv2.MORPH_ELLIPSE, (5, 5))
cv2.dilate(v, kernel, dst=v)
# Remove noise and leave cyan circle only
cv2.bitwise_and(y, v, dst=y)
# Remove game UI
cv2.bitwise_and(y, self.mask_interact, dst=y)
# Remove points on the edge, or draw_circle() will overflow
remove_border(y, max_radius)
# Get all pixels
points = inrange(y, lower=18)
# print(points.shape)
if points.shape[0] > 1000:
logger.warning(f'AimDetector.predict_item() too many points to draw: {points.shape}')
# Draw circles
draw = np.zeros((height, width), dtype=np.uint8)
draw_circle(draw, self.circle_item, points)
if self.debug:
self.draw_item = cv2.multiply(draw, 2)
# Find peaks
points = inrange(draw, lower=64)
points = Points(points).group(threshold=10)
if points.shape[0] > 3:
logger.warning(f'AimDetector.predict_item() too many peaks: {points.shape}')
self.points_item = points
# print(points)
return points
# @timer
def predict(self, image, enemy=True, item=True, show_log=True, debug=False):
"""
Predict `aim` on image, costs about 10.0~10.5ms.
Args:
image:
enemy: True to predict enemy
item: True to predict item
show_log:
debug: True to show AimDetector image
"""
self.debug = debug
self.clear_image_cache()
if isinstance(image, str):
image = load_image(image)
# 1.5~2.0ms
yuv = cv2.cvtColor(image, cv2.COLOR_RGB2YUV)
v = yuv[:, :, 2]
h = yuv[:, :, 0]
# 4.0~4.5ms
if enemy:
self.predict_enemy(h.copy(), v.copy())
# 3.0~3.5ms
if item:
self.predict_item(v.copy())
if show_log:
kv = {}
if self.aimed_enemy:
kv['enemy'] = self.aimed_enemy
if self.aimed_item:
kv['item'] = self.aimed_item
if kv:
logger.info(f'Aimed: {dict_to_kv(kv)}')
if debug:
self.show_aim()
def show_aim(self):
if self.draw_enemy is None:
if self.draw_item is None:
return
else:
r = g = b = self.draw_item
else:
if self.draw_item is None:
r = g = b = self.draw_enemy
else:
r = self.draw_enemy
g = b = self.draw_item
image = cv2.merge([b, g, r])
cv2.imshow('AimDetector', image)
cv2.waitKey(1)
@cached_property
def aimed_enemy(self) -> tuple[int, int] | None:
if self.points_enemy is None:
return None
try:
_ = self.points_enemy[1]
logger.warning(f'Multiple aimed enemy found, using first point of {self.points_enemy}')
except IndexError:
pass
try:
point = self.points_enemy[0]
return tuple(point)
except IndexError:
return None
@cached_property
def aimed_item(self) -> tuple[int, int] | None:
if self.points_item is None:
return None
try:
_ = self.points_item[1]
logger.warning(f'Multiple aimed item found, using first point of {self.points_item}')
except IndexError:
pass
try:
point = self.points_item[0]
return tuple(point)
except IndexError:
return None
class AimDetectorMixin(UI):
@cached_property
def aim(self):
return Aim()
if __name__ == '__main__':
"""
Test
"""
self = AimDetectorMixin('src')
self.device.disable_stuck_detection()
while 1:
self.device.screenshot()
self.aim.predict(self.device.image, debug=True)