SIFT, orthogonal ship image detection (lives), colored debug output

Set up a SIFT method and massively upgraded the debug view to look at the image processing outputs in color.
TODO: the template matching used for unrotated ship icons and asteroids does not work on missiles.
This commit is contained in:
John McCardle 2021-12-22 20:29:36 -05:00
parent f497bdd61a
commit 372f250167
3 changed files with 150 additions and 28 deletions

View File

@ -2,9 +2,8 @@ import gameio
import cv2
import numpy as np
def squared_distance(vec1, vec2):
"""returns distance-squared between two x, y point tuples"""
return (vec1[0] - vec2[0])**2 + (vec1[1] - vec2[1])**2
from utility import *
import pointcluster
class GameModel:
"""Platform-independent representation of the game's state."""
@ -19,19 +18,21 @@ class GameModel:
("ship_off", cv2.imread("images/game_assets/spaceship-off.png", 0)),
("ship_on", cv2.imread("images/game_assets/spaceship-on.png", 0))
]
#self.missile = ("missile", cv2.imread("images/game_assets/missile.png", 0))
self.frame = None
self.cv_template_thresh = 0.6 # reconfigurable at runtime
self.duplicate_dist_thresh = 10
self.duplicate_dist_thresh = 36
def with_frame(fn):
"""Decorator to process screenshot to cv2 format once upon first requirement, then reuse."""
def inner(self, *args, **kwargs):
if self.frame is None:
print("Fetching frame.")
#print("Fetching frame.")
sshot = self.gameio.fetch_sshot()
open_cv_image = np.array(sshot)
# Convert RGB to BGR
self.frame = open_cv_image[:, :, ::-1].copy()
self.color_frame = np.copy(self.frame)
self.frame = cv2.cvtColor(self.frame, cv2.COLOR_BGR2GRAY)
return fn(self, *args, **kwargs)
return inner
@ -52,32 +53,80 @@ class GameModel:
return asteroid_rects
@with_frame
def display_results(self, results):
def display_results(self, rects = [], pointsets = [], circles = []):
"""Draws results on the current frame for test purposes."""
displayable = np.copy(self.frame)
for pt, wh, label in results:
cv2.rectangle(displayable, pt, wh, 255, 1)
displayable = np.copy(self.color_frame)
for pt, wh, label in rects:
color = { "big": (255, 0, 0),
"normal": (0, 255, 0),
"small": (0, 0, 255),
"missile": (128, 0, 0),
"ship_on": (0, 0, 128),
"ship_off": (0, 64, 128)}[label]
cv2.rectangle(displayable, pt, wh, color, 1)
cv2.putText(displayable, label, pt,
cv2.FONT_HERSHEY_PLAIN,
1.0, 255)
1.0, color)
for ps in pointsets:
color = (0, 255, 255)
cv2.polylines(displayable, np.int32([ps]), True, color)
for center, radius, label in circles:
color = (255, 255, 0)
cv2.circle(displayable, np.int32(center), int(radius), color, 1)
cv2.putText(displayable, label, np.int32(center),
cv2.FONT_HERSHEY_PLAIN,
1.0, color)
cv2.imshow("Results", displayable)
cv2.waitKey(0)
@with_frame
def find_ships(self):
def frame_sift(self):
sift = cv2.SIFT_create()
frame_kp, frame_desc = sift.detectAndCompute(self.frame, None)
kp_desc = [] # list of (keypoints, descriptions) for all ship sprites
for label, s in self.ships:
kp_desc.append((label, sift.detectAndCompute(s, None)))
bf = cv2.BFMatcher(cv2.NORM_L1, crossCheck=True)
matchsets = []
for label, kpdesc in kp_desc:
_, desc = kpdesc
matchsets.append((label, bf.match(frame_desc, desc)))
return { "matchsets": matchsets,
"kp_desc": kp_desc
}
kp_desc = {} # dict of (keypoints, descriptions) for all ship sprites
kp_desc["frame"] = sift.detectAndCompute(self.frame, None)
frame_kp, frame_desc = kp_desc["frame"]
## for label, s in self.ships:
## kp_desc[label] = sift.detectAndCompute(s, None)
## bf = cv2.BFMatcher(cv2.NORM_L1, crossCheck=True)
## matchsets = {}
## for label in kp_desc:
## _, desc = kp_desc[label]
## matchsets[label] = bf.match(frame_desc, desc)
## #return { "matchsets": matchsets,
## # "kp_desc": kp_desc
## # }
ship_rsq = rect_radius_squared(*self.ships[0][1].shape)
#print(f"max radius^2: {ship_rsq}")
clusters = pointcluster.cluster_set([k.pt for k in frame_kp], sqrt(ship_rsq))
return clusters
@with_frame
def find_ships(self):
ship_rects = []
for label, a in self.ships:
h, w = a.shape
res = cv2.matchTemplate(self.frame, a, cv2.TM_CCOEFF_NORMED)
loc = np.where( res >= self.cv_template_thresh)
for pt in zip(*loc[::-1]):
if not ship_rects or squared_distance(ship_rects[-1][0], pt) > self.duplicate_dist_thresh:
ship_rects.append((pt, (pt[0] + w, pt[1] + h), label))
return ship_rects
## @with_frame
## def find_missiles(self):
## """This technique does not work for the 9x9 pixel missile image."""
## missile_rects = []
## label, img = self.missile
## h, w = img.shape
## res = cv2.matchTemplate(self.frame, img, cv2.TM_CCOEFF_NORMED)
## loc = np.where( res >= self.cv_template_thresh)
## for pt in zip(*loc[::-1]):
## if not missile_rects or squared_distance(missile_rects[-1][0], pt) > self.duplicate_dist_thresh:
## missile_rects.append((pt, (pt[0] + w, pt[1] + h), label))
## return missile_rects
if __name__ == '__main__':
import platform
@ -99,8 +148,14 @@ if __name__ == '__main__':
io.loc = pyscreeze.Box(0, 25, 800, 599)
#input("Press <enter> to detect asteroids on screen.")
results = gm.find_asteroids()
print(f"Found {len(results)} asteroids")
for a in results:
print(a[0]) # position tuple
gm.display_results(results)
a_results = gm.find_asteroids()
print(f"Found {len(a_results)} asteroids")
#for a in a_results:
# print(a[0]) # position tuple
#gm.display_results(results)
s_results = gm.frame_sift()
ship_results = gm.find_ships()
polygons = [c.points for c in s_results]
#circles = [(c.center, c.max_distance, f"cluster_{i}") for i, c in enumerate(s_results)]
r_circles = [(c.center, sqrt(rect_radius_squared(*gm.ships[0][1].shape)), f"cluster_{i}") for i, c in enumerate(s_results)]
gm.display_results(rects=a_results+ship_results, pointsets=polygons, circles=r_circles)

58
pointcluster.py Normal file
View File

@ -0,0 +1,58 @@
from utility import *
class PointCluster:
def __init__(self):
self.points = []
self.center = (0, 0)
self.max_distance = None
def update(self):
if len(self.points) == 0: return
self.center = (sum([p[0] for p in self.points]) / len(self.points),
sum([p[1] for p in self.points]) / len(self.points))
self.max_distance = sqrt(max(
[squared_distance(self.center, p) for p in self.points]))
def add(self, pt):
self.points.append(pt)
self.update()
def pop(self):
p = self.points.pop(-1)
self.update()
return p
def __repr__(self):
c = f"({self.center[0]:.1f},{self.center[1]:.1f})"
return f"<PointCluster center={c}, {len(self.points)} points>"
def cluster_set(points, maxradius):
"""returns a list of PointCluster objects. Points are fit within circles of maxradius"""
clusters = []
for pt in points:
if len(clusters) == 0:
#print("first cluster")
clusters.append(PointCluster())
clusters[-1].add(pt)
continue
# add point to its nearest cluster
scored_clusters = [(c, squared_distance(pt, c.center)) for c in clusters]
scored_clusters.sort(key=lambda i: i[1])
winner = scored_clusters[0][0]
winner.add(pt)
# if maxradius constraint was violated, pop the newest point & add new cluster
if winner.max_distance > maxradius:
#print(f"{winner.max_distance} > {maxradius}; new cluster")
winner.pop()
clusters.append(PointCluster())
clusters[-1].add(pt)
# refine step - accept centers as fixed, put points in closest center
new_clusters = {c.center: PointCluster() for c in clusters}
closest = lambda pt: sorted(new_clusters.keys(), key= lambda i: squared_distance(pt, i))[0]
for point in points:
new_clusters[closest(point)].add(point)
#print(clusters)
#print(new_clusters.values())
return new_clusters.values()

9
utility.py Normal file
View File

@ -0,0 +1,9 @@
from math import sqrt
def squared_distance(vec1, vec2):
"""returns distance-squared between two x, y point tuples"""
return (vec1[0] - vec2[0])**2 + (vec1[1] - vec2[1])**2
def rect_radius_squared(w, h):
"""Returns the radius^2 of the circle inscribed in a rectangle of w * h"""
return (w/2)**2 + (h/2)**2