mosse.py 6.13 KB
Newer Older
wester committed
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
#!/usr/bin/env python

'''
MOSSE tracking sample

This sample implements correlation-based tracking approach, described in [1].

Usage:
  mosse.py [--pause] [<video source>]

  --pause  -  Start with playback paused at the first video frame.
              Useful for tracking target selection.

  Draw rectangles around objects with a mouse to track them.

Keys:
  SPACE    - pause video
  c        - clear targets

[1] David S. Bolme et al. "Visual Object Tracking using Adaptive Correlation Filters"
a  
Kai Westerkamp committed
21
    http://www.cs.colostate.edu/~bolme/publications/Bolme2010Tracking.pdf
wester committed
22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170
'''

import numpy as np
import cv2
from common import draw_str, RectSelector
import video

def rnd_warp(a):
    h, w = a.shape[:2]
    T = np.zeros((2, 3))
    coef = 0.2
    ang = (np.random.rand()-0.5)*coef
    c, s = np.cos(ang), np.sin(ang)
    T[:2, :2] = [[c,-s], [s, c]]
    T[:2, :2] += (np.random.rand(2, 2) - 0.5)*coef
    c = (w/2, h/2)
    T[:,2] = c - np.dot(T[:2, :2], c)
    return cv2.warpAffine(a, T, (w, h), borderMode = cv2.BORDER_REFLECT)

def divSpec(A, B):
    Ar, Ai = A[...,0], A[...,1]
    Br, Bi = B[...,0], B[...,1]
    C = (Ar+1j*Ai)/(Br+1j*Bi)
    C = np.dstack([np.real(C), np.imag(C)]).copy()
    return C

eps = 1e-5

class MOSSE:
    def __init__(self, frame, rect):
        x1, y1, x2, y2 = rect
        w, h = map(cv2.getOptimalDFTSize, [x2-x1, y2-y1])
        x1, y1 = (x1+x2-w)//2, (y1+y2-h)//2
        self.pos = x, y = x1+0.5*(w-1), y1+0.5*(h-1)
        self.size = w, h
        img = cv2.getRectSubPix(frame, (w, h), (x, y))

        self.win = cv2.createHanningWindow((w, h), cv2.CV_32F)
        g = np.zeros((h, w), np.float32)
        g[h//2, w//2] = 1
        g = cv2.GaussianBlur(g, (-1, -1), 2.0)
        g /= g.max()

        self.G = cv2.dft(g, flags=cv2.DFT_COMPLEX_OUTPUT)
        self.H1 = np.zeros_like(self.G)
        self.H2 = np.zeros_like(self.G)
        for i in xrange(128):
            a = self.preprocess(rnd_warp(img))
            A = cv2.dft(a, flags=cv2.DFT_COMPLEX_OUTPUT)
            self.H1 += cv2.mulSpectrums(self.G, A, 0, conjB=True)
            self.H2 += cv2.mulSpectrums(     A, A, 0, conjB=True)
        self.update_kernel()
        self.update(frame)

    def update(self, frame, rate = 0.125):
        (x, y), (w, h) = self.pos, self.size
        self.last_img = img = cv2.getRectSubPix(frame, (w, h), (x, y))
        img = self.preprocess(img)
        self.last_resp, (dx, dy), self.psr = self.correlate(img)
        self.good = self.psr > 8.0
        if not self.good:
            return

        self.pos = x+dx, y+dy
        self.last_img = img = cv2.getRectSubPix(frame, (w, h), self.pos)
        img = self.preprocess(img)

        A = cv2.dft(img, flags=cv2.DFT_COMPLEX_OUTPUT)
        H1 = cv2.mulSpectrums(self.G, A, 0, conjB=True)
        H2 = cv2.mulSpectrums(     A, A, 0, conjB=True)
        self.H1 = self.H1 * (1.0-rate) + H1 * rate
        self.H2 = self.H2 * (1.0-rate) + H2 * rate
        self.update_kernel()

    @property
    def state_vis(self):
        f = cv2.idft(self.H, flags=cv2.DFT_SCALE | cv2.DFT_REAL_OUTPUT )
        h, w = f.shape
        f = np.roll(f, -h//2, 0)
        f = np.roll(f, -w//2, 1)
        kernel = np.uint8( (f-f.min()) / f.ptp()*255 )
        resp = self.last_resp
        resp = np.uint8(np.clip(resp/resp.max(), 0, 1)*255)
        vis = np.hstack([self.last_img, kernel, resp])
        return vis

    def draw_state(self, vis):
        (x, y), (w, h) = self.pos, self.size
        x1, y1, x2, y2 = int(x-0.5*w), int(y-0.5*h), int(x+0.5*w), int(y+0.5*h)
        cv2.rectangle(vis, (x1, y1), (x2, y2), (0, 0, 255))
        if self.good:
            cv2.circle(vis, (int(x), int(y)), 2, (0, 0, 255), -1)
        else:
            cv2.line(vis, (x1, y1), (x2, y2), (0, 0, 255))
            cv2.line(vis, (x2, y1), (x1, y2), (0, 0, 255))
        draw_str(vis, (x1, y2+16), 'PSR: %.2f' % self.psr)

    def preprocess(self, img):
        img = np.log(np.float32(img)+1.0)
        img = (img-img.mean()) / (img.std()+eps)
        return img*self.win

    def correlate(self, img):
        C = cv2.mulSpectrums(cv2.dft(img, flags=cv2.DFT_COMPLEX_OUTPUT), self.H, 0, conjB=True)
        resp = cv2.idft(C, flags=cv2.DFT_SCALE | cv2.DFT_REAL_OUTPUT)
        h, w = resp.shape
        _, mval, _, (mx, my) = cv2.minMaxLoc(resp)
        side_resp = resp.copy()
        cv2.rectangle(side_resp, (mx-5, my-5), (mx+5, my+5), 0, -1)
        smean, sstd = side_resp.mean(), side_resp.std()
        psr = (mval-smean) / (sstd+eps)
        return resp, (mx-w//2, my-h//2), psr

    def update_kernel(self):
        self.H = divSpec(self.H1, self.H2)
        self.H[...,1] *= -1

class App:
    def __init__(self, video_src, paused = False):
        self.cap = video.create_capture(video_src)
        _, self.frame = self.cap.read()
        cv2.imshow('frame', self.frame)
        self.rect_sel = RectSelector('frame', self.onrect)
        self.trackers = []
        self.paused = paused

    def onrect(self, rect):
        frame_gray = cv2.cvtColor(self.frame, cv2.COLOR_BGR2GRAY)
        tracker = MOSSE(frame_gray, rect)
        self.trackers.append(tracker)

    def run(self):
        while True:
            if not self.paused:
                ret, self.frame = self.cap.read()
                if not ret:
                    break
                frame_gray = cv2.cvtColor(self.frame, cv2.COLOR_BGR2GRAY)
                for tracker in self.trackers:
                    tracker.update(frame_gray)

            vis = self.frame.copy()
            for tracker in self.trackers:
                tracker.draw_state(vis)
            if len(self.trackers) > 0:
                cv2.imshow('tracker state', self.trackers[-1].state_vis)
            self.rect_sel.draw(vis)

            cv2.imshow('frame', vis)
wester committed
171
            ch = cv2.waitKey(10)
wester committed
172 173 174 175 176 177 178 179 180
            if ch == 27:
                break
            if ch == ord(' '):
                self.paused = not self.paused
            if ch == ord('c'):
                self.trackers = []


if __name__ == '__main__':
wester committed
181
    print __doc__
wester committed
182 183 184
    import sys, getopt
    opts, args = getopt.getopt(sys.argv[1:], '', ['pause'])
    opts = dict(opts)
wester committed
185 186
    try: video_src = args[0]
    except: video_src = '0'
wester committed
187 188

    App(video_src, paused = '--pause' in opts).run()