1. fix improper size filter; 2. update a grading func; 3. clear fig win (df06ae49) · Commits · Ziqin Wang / ArmorDetection

src/car_target.py

+5 −5

Original line number	Diff line number	Diff line
		#!/usr/bin/env python3
		# -- coding: utf-8 --
		__date__ = '2018/04/06'

		import cv2
		import math
		@@ -77,11 +76,12 @@ class ImgCarTargetApp(CarTargetApp):
		@quitable
		def run(self): # FIXME: move folder path to constructor
		file_list = folder(self.folder, self.ext_name)
		plt.figure("Armor Detection")
		for i in file_list:
		print("img_file:", i)
		mat = cv2.resize(cv2.imread(i), self.frame_size)

		plt.figure("Armor Detection")
		plt.clf()

		ROW, COL = 2, 4
		count = 1
		@@ -219,11 +219,11 @@ if __name__ == "__main__":

		# app = ImgCarTargetApp(color=BarColor.BLUE, folder="/home/jeeken/Pictures/blue",
		# frame_size=(640, 480), ext_name="jpg", debug=True)
		# app = ImgCarTargetApp(color=BarColor.RED, folder="/home/jeeken/Pictures/red",
		# frame_size=(640, 360), ext_name="jpg", debug=True)
		# app = ImgCarTargetApp(color=BarColor.RED, folder="/home/jeeken/Pictures/red_dark",
		# frame_size=(640, 480), ext_name="jpg", debug=True)

		app = VideoCarTargetApp(file="/home/jeeken/Videos/live_blue.avi",
		color=BarColor.BLUE, frame_size=(640, 480), debug=False)
		# app = VideoCarTargetApp(file="/home/jeeken/Videos/red.avi",
		# color=BarColor.RED, frame_size=(640, 480), debug=True)
		# color=BarColor.RED, frame_size=(640, 480), debug=False)
		app.run()

src/detector.py

+57 −27

Original line number	Diff line number	Diff line
		@@ -81,25 +81,38 @@ class Detector:
		upper_light = np.array([120, 130, 255])
		lower_halo = np.array([90, 100, 185])
		upper_halo = np.array([120, 255, 255])
		# lower_light = np.array([90, 0, 185])
		# upper_light = np.array([120, 230, 255])
		# lower_halo = np.array([90, 30, 160])
		# upper_halo = np.array([120, 255, 255])

		light_area = cv2.inRange(hsv_frame, lower_light, upper_light)
		halo_area = cv2.inRange(hsv_frame, lower_halo, upper_halo)
		return light_area, halo_area

		def get_lights(self, binary_img: np.ndarray):
		_, width = binary_img.shape
		# https://docs.opencv.org/2.4/modules/imgproc/doc/structural_analysis_and_shape_descriptors.html?highlight=findcontours#findcontours
		img, contours, hierarchy = cv2.findContours(binary_img, mode=cv2.RETR_EXTERNAL, method=cv2.CHAIN_APPROX_SIMPLE)
		min_width = width // 60
		rectangles_info = [cv2.minAreaRect(i) for i in contours if len(i) > min_width] # filter (omit small ones) and map
		# rectangles_info = [cv2.fitEllipse(i) for i in contours if len(i) > min_width]
		return rectangles_info

		very_small_area = binary_img.shape[1] / 30 # img_width / 30

		def not_very_small(rec):
		w, h = rec[1]
		return w * h > very_small_area
		return filter(not_very_small, map(cv2.minAreaRect, contours))

		def halo_circle(self, binary_img: np.ndarray):
		height, width = binary_img.shape
		img, contours, hierarchy = cv2.findContours(binary_img, mode=cv2.RETR_EXTERNAL, method=cv2.CHAIN_APPROX_SIMPLE)
		min_width = width // 60
		selected_contours = [i for i in contours if len(i) > min_width] # small contours omitted

		very_small_area = width / 30
		def not_very_small(contour):
		points_x, points_y = cv2.split(contour)
		w = np.amax(points_x) - np.amin(points_x)
		h = np.amax(points_y) - np.amin(points_y)
		return w * h > very_small_area
		selected_contours = [i for i in contours if not_very_small(i)]

		if len(selected_contours) > 0:
		combined_halo_points = np.concatenate(selected_contours)
		original_center, original_radius = cv2.minEnclosingCircle(combined_halo_points)
		@@ -123,12 +136,23 @@ class Detector:
		def is_inside(light):
		(xl, yl), shape, angle = light
		return (xl-x)2 + (yl-y)2 < halo_radius**2
		return filter(is_inside, lights)
		# return filter(is_inside, lights)
		lights_inside = list(filter(is_inside, lights))

		# debug
		def get_img_lights_inside():
		mat = np.copy(self.frame)
		for center, shape, angle in lights_inside:
		cv2.ellipse(mat, center=round_point(center), axes=round_point(shape), angle=angle,
		startAngle=0, endAngle=360, color=BGR_GREEN, thickness=2)
		return mat
		self.add_debug_img("all lights inside", get_img_lights_inside)

		return lights_inside

		def select_vertical_lights(self, lights):
		def normalize_angle(rec):
		center, shape, angle = rec
		w, h = shape
		center, (w, h), angle = rec
		if w > h:
		return center, (h, w), angle+90
		else:
		@@ -136,7 +160,7 @@ class Detector:

		def is_vertical(rec):
		center, shape, angle = rec
		return -20 < angle < 20
		return -25 < angle < 25

		angle_normalized_lights = map(normalize_angle, lights)
		vertical_lights = [i for i in angle_normalized_lights if is_vertical(i)]
		@@ -148,7 +172,7 @@ class Detector:
		center, shape, angle = i
		rand_color = (randint(40, 230), randint(40, 255), randint(40, 255))
		cv2.ellipse(show_lights, center=round_point(center), axes=round_point(shape), angle=angle,
		startAngle=0, endAngle=360, color=rand_color, thickness=2)
		startAngle=0, endAngle=360, color=BGR_GREEN, thickness=2)
		return show_lights
		self.add_debug_img("Vertical Lights", get_img_vertical_lights)
		return vertical_lights
		@@ -164,19 +188,23 @@ class Detector:

		def parallel(light1, light2):
		angle1, angle2 = light1[2], light2[2]
		return abs(angle1 - angle2) / 20
		return abs(angle1 - angle2) / 15

		def size_similarity(light1, light2):
		def shape_similarity(light1, light2):
		(w1, h1), (w2, h2) = light1[1], light2[1]
		area1, area2 = w1 * h1, w2 * h2
		min_size = area1 if area1 < area2 else area2
		return abs(area1 - area2) / min_size
		min_width = np.minimum(w1, w2)
		min_height = np.minimum(h1, h2)
		return abs(w1-w2)/min_width + 0.5*abs(h1-h2)/min_height

		# area1, area2 = w1 * h1, w2 * h2
		# min_size = area1 if area1 < area2 else area2
		# return abs(area1 - area2) / min_size

		def square_ratio(light1, light2):
		(x1, y1), (w1, h1), angle1 = light1
		(x2, y2), (w2, h2), angle2 = light2
		ratio = 2 * math.sqrt((x1-x2)2 + (y1-y2)2) / (h1+h2)
		return abs(ratio - 2.4) ** 2 if ratio > 0.85 else 1e9
		return abs(ratio - 2.5) ** 2 if ratio > 0.85 else 1e9

		def y_dis(light1, light2):
		(x1, y1), shape1, angle1 = light1
		@@ -187,29 +215,30 @@ class Detector:
		def judge(pair):
		i1, i2 = pair
		bar1, bar2 = lights[i1], lights[i2]
		policy = [(square_ratio, 5), (y_dis, 18), (size_similarity, 1), (parallel, 1.2)] # [(func, coefficient),...]
		return sum((func(bar1, bar2) * coefficient for func, coefficient in policy)) # weighted sum
		policy = [(square_ratio, 5), (y_dis, 20), (shape_similarity, 3), (parallel, 1.2)] # [(func, coefficient),...]
		return sum([func(bar1, bar2) * coefficient for func, coefficient in policy]) # weighted sum

		judge_results = [(pair, judge(pair)) for pair in pairs]
		(i1, i2), winner_score = min(judge_results, key=lambda x: x[1])

		if winner_score > 5:
		if winner_score > 7:
		return None

		# debug
		def get_img_selected_pair():
		selected_pair_show = np.copy(self.frame)
		l1, l2 = lights[i1], lights[i2]
		# self.debug_print("l1:", l1)
		# self.debug_print("l2:", l2)
		# self.debug_print("score:", winner_score)
		# self.debug_print("parallel:", 1.2 * parallel(l1, l2))
		# self.debug_print("size_similarity:", size_similarity(l1, l2))
		# self.debug_print("shape_similarity:", 5 * size_similarity(l1, l2))
		# self.debug_print("square_ratio:", 5 * square_ratio(l1, l2))
		# self.debug_print("y_dis:", 18 * y_dis(l1, l2))
		# self.debug_print("y_dis:", 20 * y_dis(l1, l2))
		self.debug_print("score:", winner_score)
		c1, s1, a1 = l1
		c2, s2, a2 = l2
		cv2.drawMarker(selected_pair_show, position=round_point(c1), color=BGR_YELLOW, markerSize=15, thickness=2)
		cv2.drawMarker(selected_pair_show, position=round_point(c2), color=BGR_YELLOW, markerSize=15, thickness=2)
		cv2.drawMarker(selected_pair_show, position=round_point(c1), color=BGR_GREEN, markerSize=15, thickness=2)
		cv2.drawMarker(selected_pair_show, position=round_point(c2), color=BGR_GREEN, markerSize=15, thickness=2)
		return selected_pair_show
		self.add_debug_img("Selected Pair", get_img_selected_pair)

		@@ -227,6 +256,7 @@ class Detector:
		halo_center, halo_radius = self.halo_circle(halo)
		lights_info = self.get_lights(light)

		# debug
		# only draw on light/halo after info got, or draw on a copied mat
		def get_img_halo():
		cv2.circle(img=halo, center=halo_center, radius=halo_radius, color=255, thickness=2)

Admin message