1#!/usr/bin/env python 2 3''' 4Feature homography 5================== 6 7Example of using features2d framework for interactive video homography matching. 8ORB features and FLANN matcher are used. The actual tracking is implemented by 9PlaneTracker class in plane_tracker.py 10''' 11 12# Python 2/3 compatibility 13from __future__ import print_function 14 15import numpy as np 16import cv2 as cv 17import sys 18PY3 = sys.version_info[0] == 3 19 20if PY3: 21 xrange = range 22 23# local modules 24from tst_scene_render import TestSceneRender 25 26def intersectionRate(s1, s2): 27 28 x1, y1, x2, y2 = s1 29 s1 = np.array([[x1, y1], [x2,y1], [x2, y2], [x1, y2]]) 30 31 area, _intersection = cv.intersectConvexConvex(s1, np.array(s2)) 32 return 2 * area / (cv.contourArea(s1) + cv.contourArea(np.array(s2))) 33 34from tests_common import NewOpenCVTests 35 36class feature_homography_test(NewOpenCVTests): 37 38 render = None 39 tracker = None 40 framesCounter = 0 41 frame = None 42 43 def test_feature_homography(self): 44 45 self.render = TestSceneRender(self.get_sample('samples/data/graf1.png'), 46 self.get_sample('samples/data/box.png'), noise = 0.5, speed = 0.5) 47 self.frame = self.render.getNextFrame() 48 self.tracker = PlaneTracker() 49 self.tracker.clear() 50 self.tracker.add_target(self.frame, self.render.getCurrentRect()) 51 52 while self.framesCounter < 100: 53 self.framesCounter += 1 54 tracked = self.tracker.track(self.frame) 55 if len(tracked) > 0: 56 tracked = tracked[0] 57 self.assertGreater(intersectionRate(self.render.getCurrentRect(), np.int32(tracked.quad)), 0.6) 58 else: 59 self.assertEqual(0, 1, 'Tracking error') 60 self.frame = self.render.getNextFrame() 61 62 63# built-in modules 64from collections import namedtuple 65 66FLANN_INDEX_KDTREE = 1 67FLANN_INDEX_LSH = 6 68flann_params= dict(algorithm = FLANN_INDEX_LSH, 69 table_number = 6, # 12 70 key_size = 12, # 20 71 multi_probe_level = 1) #2 72 73MIN_MATCH_COUNT = 10 74 75''' 76 image - image to track 77 rect - tracked rectangle (x1, y1, x2, y2) 78 keypoints - keypoints detected inside rect 79 descrs - their descriptors 80 data - some user-provided data 81''' 82PlanarTarget = namedtuple('PlaneTarget', 'image, rect, keypoints, descrs, data') 83 84''' 85 target - reference to PlanarTarget 86 p0 - matched points coords in target image 87 p1 - matched points coords in input frame 88 H - homography matrix from p0 to p1 89 quad - target boundary quad in input frame 90''' 91TrackedTarget = namedtuple('TrackedTarget', 'target, p0, p1, H, quad') 92 93class PlaneTracker: 94 def __init__(self): 95 self.detector = cv.AKAZE_create(threshold = 0.003) 96 self.matcher = cv.FlannBasedMatcher(flann_params, {}) # bug : need to pass empty dict (#1329) 97 self.targets = [] 98 self.frame_points = [] 99 100 def add_target(self, image, rect, data=None): 101 '''Add a new tracking target.''' 102 x0, y0, x1, y1 = rect 103 raw_points, raw_descrs = self.detect_features(image) 104 points, descs = [], [] 105 for kp, desc in zip(raw_points, raw_descrs): 106 x, y = kp.pt 107 if x0 <= x <= x1 and y0 <= y <= y1: 108 points.append(kp) 109 descs.append(desc) 110 descs = np.uint8(descs) 111 self.matcher.add([descs]) 112 target = PlanarTarget(image = image, rect=rect, keypoints = points, descrs=descs, data=data) 113 self.targets.append(target) 114 115 def clear(self): 116 '''Remove all targets''' 117 self.targets = [] 118 self.matcher.clear() 119 120 def track(self, frame): 121 '''Returns a list of detected TrackedTarget objects''' 122 self.frame_points, frame_descrs = self.detect_features(frame) 123 if len(self.frame_points) < MIN_MATCH_COUNT: 124 return [] 125 matches = self.matcher.knnMatch(frame_descrs, k = 2) 126 matches = [m[0] for m in matches if len(m) == 2 and m[0].distance < m[1].distance * 0.75] 127 if len(matches) < MIN_MATCH_COUNT: 128 return [] 129 matches_by_id = [[] for _ in xrange(len(self.targets))] 130 for m in matches: 131 matches_by_id[m.imgIdx].append(m) 132 tracked = [] 133 for imgIdx, matches in enumerate(matches_by_id): 134 if len(matches) < MIN_MATCH_COUNT: 135 continue 136 target = self.targets[imgIdx] 137 p0 = [target.keypoints[m.trainIdx].pt for m in matches] 138 p1 = [self.frame_points[m.queryIdx].pt for m in matches] 139 p0, p1 = np.float32((p0, p1)) 140 H, status = cv.findHomography(p0, p1, cv.RANSAC, 3.0) 141 status = status.ravel() != 0 142 if status.sum() < MIN_MATCH_COUNT: 143 continue 144 p0, p1 = p0[status], p1[status] 145 146 x0, y0, x1, y1 = target.rect 147 quad = np.float32([[x0, y0], [x1, y0], [x1, y1], [x0, y1]]) 148 quad = cv.perspectiveTransform(quad.reshape(1, -1, 2), H).reshape(-1, 2) 149 150 track = TrackedTarget(target=target, p0=p0, p1=p1, H=H, quad=quad) 151 tracked.append(track) 152 tracked.sort(key = lambda t: len(t.p0), reverse=True) 153 return tracked 154 155 def detect_features(self, frame): 156 '''detect_features(self, frame) -> keypoints, descrs''' 157 keypoints, descrs = self.detector.detectAndCompute(frame, None) 158 if descrs is None: # detectAndCompute returns descs=None if no keypoints found 159 descrs = [] 160 return keypoints, descrs 161 162 163if __name__ == '__main__': 164 NewOpenCVTests.bootstrap() 165