- coarse/fine angle estimation

- speed up angle estimation

ocv_track.py

@@ -4,7 +4,7 @@ import numpy as np
 import imutils as im
 
 IMG_SCALE_UP = 4
-TEMPLATE_SEARCH_AREA = 0.5
+TEMPLATE_SEARCH_AREA = 0.8
 
 (major_ver, minor_ver, subminor_ver) = cv2.__version__.split('.')
 print(cv2.__version__)
@@ -21,8 +21,19 @@ def bbox_extend(bbox: cv2.typing.Rect, search_area: float):
 	return res
 
 
+def bbox_round(src):
+	x = int(round(src[0]))
+	y = int(round(src[1]))
+	w = int(round(src[2]))
+	h = int(round(src[3]))
+
+	return x, y, w, h
+
+
 def image_crop(src, bbox: cv2.typing.Rect, search_area: float = 0):
-	bbox = bbox_extend(bbox, search_area)
+	if search_area > 0:
+		bbox = bbox_extend(bbox, search_area)
+
 	x = bbox[0]
 	y = bbox[1]
 	w = bbox[2]
@@ -39,16 +50,15 @@ def process_image_bbox(src):
 	return gray
 
 
-def match(source, template, angle, center):
+def template_match_rotation(source, template, angle, center, rot_min=-1.0, rot_max=+1.0, n_steps=10, scale=1):
 	best_angle = 0
 	best_r = 0
 	best_max_loc = 0
-
-	im_source = im.resize(source, IMG_SCALE_UP * source.shape[0], IMG_SCALE_UP * source.shape[1])
-
-	for dangle in np.linspace(-1.0, +1.0, 10):
+	r = 0
+	im_source = im.resize(source, scale * source.shape[0], scale * source.shape[1])
+	for dangle in np.linspace(rot_min, rot_max, n_steps):
 		# rotated and scale up to IMG_SCALE_UP-size
-		im_template = im.rotate(template, angle + dangle, center=center, scale=IMG_SCALE_UP)
+		im_template = im.rotate(template, angle + dangle, center=center, scale=scale)
 
 		# Perform template match operations
 		res = cv2.matchTemplate(im_source, im_template, cv2.TM_CCOEFF_NORMED)
@@ -58,13 +68,24 @@ def match(source, template, angle, center):
 		if r > best_r:
 			best_r = r
 			best_angle = dangle
-			best_max_loc = (maxLoc[0]/IMG_SCALE_UP, maxLoc[1]/IMG_SCALE_UP)
+			best_max_loc = (maxLoc[0]/scale, maxLoc[1]/scale)
+		else:
+			break
 
 	angle += best_angle
 
 	return angle, best_max_loc
 
 
+def match(source, template, angle, center):
+	# coarse
+	angle_c, loc = template_match_rotation(source, template, angle, center)
+	# fine
+	angle, loc = template_match_rotation(source, template, angle_c, center, rot_min=-0.25, rot_max=+0.25, n_steps=20, scale=IMG_SCALE_UP)
+
+	return angle, loc
+
+
 class Reference:
 	def __init__(self, ref_gray):
 		self.center = (0, 0)
@@ -104,7 +125,7 @@ trackers = {
 	'BOOSTING': cv2.legacy.TrackerBoosting,
 	'MEDIANFLOW': cv2.legacy.TrackerMedianFlow
 }
-tracker_type = 'MOSSE'
+tracker_type = 'KCF'
 tracker = trackers[tracker_type].create()
 
 video = cv2.VideoCapture("./data/production_id 4525346 (1080p).mp4")
@@ -142,18 +163,19 @@ while True:
 	timer = cv2.getTickCount()
 
 	# Update tracker
-	ok, bbox = tracker.update(frame)
+	ok, bbox_tracker = tracker.update(frame)
 
-	image_bbox = process_image_bbox(image_crop(frame, bbox))
 	image_template_rotated = image_template
+	bbox = bbox_round(bbox_tracker)
 
 	if ok:
+		image_bbox = process_image_bbox(image_crop(frame, bbox))
 		current_angle, max_loc = match(image_bbox, image_template, current_angle, ref_center)
 
 		# determine the starting and ending (x, y)-coordinates of the
 		# bounding box
 		(startX, startY) = max_loc
-		print(f"startX: {startX}, startY: {startY}")
+		print(f"current_angle: {current_angle:.2f}, startX: {startX}, startY: {startY}")
 		image_template_rotated = im.rotate(image_template, current_angle, center=ref_center)
 
 		# draw the bounding box on the image