如何检测两个PIL图像之间的运动？（包含wxPython webcams整合示例）

Question

有没有人能给我一些建议，如何在Python中进行图像比较，以检测图像中的变化？我现在正在开发一个应用程序，用我的网络摄像头监控我的区域，我想知道如何比较每一帧拍摄的图像，以查看是否检测到任何运动。长远来看，我希望能够设置一个灵敏度滑块，所以如果你能给我一些指导，我相信我能搞定剩下的部分。

我看到这里有一些帖子在讨论如何将网络摄像头与wxPython结合使用，这里有一个小示例。请注意，我昨晚才开始做这个，所以如果你在寻找完美的代码，可能需要自己修改一下（暂时如此；）:

需求： PIL 和 VideoCapture

#videocapturepanel.py

#Todo:
# - Fix background colour after video is stopped
# - Create image comparison method
# - Add capture function
# - Save stream to video file?


import threading, wx
from PIL          import Image
from VideoCapture import Device

cam = Device(0)
buffer, width, height = cam.getBuffer()
cam.setResolution(width, height)

DEFAULT_DEVICE_INDEX  = 0
DEFAULT_DEVICE_WIDTH  = width
DEFAULT_DEVICE_HEIGHT = height
DEFAULT_BACKGROUND_COLOUR = wx.Colour(0, 0, 0)

class VideoCaptureThread(threading.Thread):

    def __init__(self, control, width=DEFAULT_DEVICE_WIDTH, height=DEFAULT_DEVICE_HEIGHT, backColour=DEFAULT_BACKGROUND_COLOUR):
        self.backColour = backColour
        self.width      = width
        self.height     = height
        self.control    = control
        self.isRunning  = True
        self.buffer     = wx.NullBitmap

        threading.Thread.__init__(self)

    def getResolution(self):
        return (self.width, self.height)

    def setResolution(self, width, height):
        self.width  = width
        self.height = height
        cam.setResolution(width, height)

    def getBackgroundColour(self):
        return self.backColour

    def setBackgroundColour(self, colour):
        self.backColour = colour

    def getBuffer(self):
        return self.buffer

    def stop(self):
        self.isRunning = False

    def run(self):
        while self.isRunning:
            buffer, width, height = cam.getBuffer()
            im = Image.fromstring('RGB', (width, height), buffer, 'raw', 'BGR', 0, -1)
            buff = im.tostring()
            self.buffer = wx.BitmapFromBuffer(width, height, buff)
            x, y = (0, 0)
            try:
                width, height = self.control.GetSize()
                if width > self.width:
                    x = (width - self.width) / 2
                if height > self.height:
                    y = (height - self.height) / 2
                dc = wx.BufferedDC(wx.ClientDC(self.control), wx.NullBitmap, wx.BUFFER_VIRTUAL_AREA)
                dc.SetBackground(wx.Brush(self.backColour))
                dc.Clear()
                dc.DrawBitmap(self.buffer, x, y)
            except TypeError:
                pass
            except wx.PyDeadObjectError:
                pass
        self.isRunning = False


class VideoCapturePanel(wx.Panel):

    def __init__(self, parent, id=-1, pos=wx.DefaultPosition, size=wx.DefaultSize, initVideo=False, style=wx.SUNKEN_BORDER):
        wx.Panel.__init__(self, parent, id, pos, size, style)

        if initVideo:
            self.StartVideo()

        self.Bind(wx.EVT_CLOSE, self.OnClose)

    def OnClose(self, event):
        try:
            self.Device.stop()
        except:
            pass

    def StopVideo(self):
        self.Device.stop()
        self.SetBackgroundColour(self.Device.backColour)
        dc = wx.BufferedDC(wx.ClientDC(self), wx.NullBitmap)
        dc.SetBackground(wx.Brush(self.Device.backColour))
        dc.Clear()

    def StartVideo(self):
        self.Device = VideoCaptureThread(self)
        self.Device.start()

    def GetBackgroundColour(self):
        return self.Device.getBackgroundColour()

    def SetBackgroundColour(self, colour):
        self.Device.setBackgroundColour(colour)


class Frame(wx.Frame):

    def __init__(self, parent, id=-1, title="A Frame", path="", pos=wx.DefaultPosition, size=wx.DefaultSize, style=wx.DEFAULT_FRAME_STYLE):
        wx.Frame.__init__(self, parent, id, title, pos, size, style)
        self.VidPanel = VideoCapturePanel(self, -1, initVideo=False)
        self.StartButton  = wx.ToggleButton(self, -1, "Turn On")
        self.ColourButton = wx.Button(self, -1, "Change Background")
        szr  = wx.BoxSizer(wx.VERTICAL)
        bszr = wx.BoxSizer(wx.HORIZONTAL)
        bszr.Add(self.StartButton, 0, wx.ALIGN_CENTER_HORIZONTAL | wx.LEFT, 5)
        bszr.Add(self.ColourButton, 0, wx.ALIGN_CENTER_HORIZONTAL)
        szr.Add(self.VidPanel, 1, wx.EXPAND)
        szr.Add(bszr, 0, wx.ALIGN_CENTER_HORIZONTAL)
        self.SetSizer(szr)

        self.StartButton.Bind(wx.EVT_TOGGLEBUTTON, self.OnToggled)
        self.ColourButton.Bind(wx.EVT_BUTTON, self.OnColour)


    def OnColour(self, event):
        dlg = wx.ColourDialog(self)
        dlg.GetColourData().SetChooseFull(True)

        if dlg.ShowModal() == wx.ID_OK:
            data = dlg.GetColourData()
            self.VidPanel.SetBackgroundColour(data.GetColour())
        dlg.Destroy()


    def OnToggled(self, event):
        if event.IsChecked():
            self.VidPanel.StartVideo()
        else:
            self.VidPanel.StopVideo()
            #self.VidPanel.SetBackgroundColour(data.GetColour())


if __name__ == "__main__":
    # Run GUI
    app   = wx.PySimpleApp()
    frame = Frame(None, -1, "Test Frame", size=(800, 600))
    frame.Show()
    app.MainLoop()
    del app

*更新*

根据Paul的示例，我创建了一个类并将其整合到我的代码中：

class Images:

    def __init__(self, image1, image2, threshold=98, grayscale=True):
        self.image1 = image1
        if type(image1) == str:
            self.image1 = Image.open(self.image1)
        self.image2 = image2
        if type(image2) == str:
            self.image2 = Image.open(image2)
        self.threshold = threshold

    def DoComparison(self, image1=None, image2=None):
        if not image1: image1 = self.image1
        if not image2: image2 = self.image2
        diffs = ImageChops.difference(image1, image2)
        return self.ImageEntropy(diffs)

    def ImageEntropy(self, image):
        histogram   = image.histogram()
        histlength  = sum(histogram)
        probability = [float(h) / histlength for h in histogram]
        return -sum([p * math.log(p, 2) for p in probability if p != 0])

然后在VideoCaptureThread的__init__()函数中添加了变量self.image = False，并在VideoCaptureThread的run()函数中，在im = Image.fromstring(...)这一行之后添加了下面的代码：

        if self.image:
            img = compare.Images2(im, self.image).DoComparison()
            print img
        self.image = im

当我运行这个示例时，似乎工作得还不错，但我对得到的结果有点困惑：

1.58496250072
5.44792407663
1.58496250072
5.44302784225
1.58496250072
5.59144486002
1.58496250072
5.37568050189
1.58496250072

到目前为止，似乎每隔一张图像就会有很大的偏差，尽管变化很小？理论上，添加到run中的代码应该会捕捉到之前的图像，并将其存储在self.image变量中，然后与新图像im进行比较。比较之后，self.image会用当前图像更新，使用self.image = im，那么为什么每第二张图像会有这么大的差异呢？最多我的眼睛在两张图像之间可能有些移动，我看不出这会导致结果有如此大的差别？

*更新 2*

这是我目前的进展，有三个比较类，使用三种不同的方法来检测运动。

class Images ~ 第一次尝试使用我在网上找到的一些代码，老实说，我甚至不记得它是怎么工作的。:P

class Images2 ~ 使用Paul在这个帖子中的代码创建，实施了他更新的图像熵函数。

class Images3 ~ 修改版的DetectMotion函数，找到的链接在这里。（返回变化的百分比，并似乎考虑了光照因素）

老实说，我真的不知道它们在做什么，字面意思，但我能说的是，到目前为止，class Image3似乎是设置检测的最简单/准确的方法，缺点是它的处理时间比其他两个类要长。

(请注意，进行了一些导入更改，以避免与scipy冲突，sys.modules["Image"]与PIL.Image是相同的)

import math, sys, numpy as np
import PIL.Image, PIL.ImageChops

sys.modules["Image"]      = PIL.Image
sys.modules["ImageChops"] = PIL.ImageChops

from scipy.misc   import imread
from scipy.linalg import norm
from scipy        import sum, average


DEFAULT_DEVICE_WIDTH  = 640
DEFAULT_DEVICE_HEIGHT = 480


class Images:

    def __init__(self, image1, image2, threshold=98, grayscale=True):
        if type(image1) == str:
            self.image1 = sys.modules["Image"].open(image1)
            self.image2 = sys.modules["Image"].open(image2)
        if grayscale:
            self.image1 = self.DoGrayscale(imread(image1).astype(float))
            self.image2 = self.DoGrayscale(imread(image2).astype(float))
        else:
            self.image1    = imread(image1).astype(float)
            self.image2    = imread(image2).astype(float)
        self.threshold = threshold

    def DoComparison(self, image1=None, image2=None):
        if image1: image1 = self.Normalize(image1)
        else:      image1 = self.Normalize(self.image1)
        if image2: image2 = self.Normalize(image2)
        else:      image2 = self.Normalize(self.image2)
        diff = image1 - image2
        m_norm = sum(abs(diff))
        z_norm = norm(diff.ravel(), 0)
        return (m_norm, z_norm)

    def DoGrayscale(self, arr):
        if len(arr.shape) == 3:
            return average(arr, -1)
        else:
            return arr

    def Normalize(self, arr):
        rng = arr.max()-arr.min()
        amin = arr.min()
        return (arr-amin)*255/rng


class Images2:

    def __init__(self, image1, image2, threshold=98, grayscale=True):
        self.image1 = image1
        if type(image1) == str:
            self.image1 = sys.modules["Image"].open(self.image1)
        self.image2 = image2
        if type(image2) == str:
            self.image2 = sys.modules["Image"].open(image2)
        self.threshold = threshold

    def DoComparison(self, image1=None, image2=None):
        if not image1: image1 = self.image1
        if not image2: image2 = self.image2
        diffs = sys.modules["ImageChops"].difference(image1, image2)
        return self.ImageEntropy(diffs)

    def ImageEntropy(self, image):
        w,h = image.size
        a = np.array(image.convert('RGB')).reshape((w*h,3))
        h,e = np.histogramdd(a, bins=(16,)*3, range=((0,256),)*3)
        prob = h/np.sum(h)
        return -np.sum(np.log2(prob[prob>0]))


    def OldImageEntropy(self, image):
        histogram   = image.histogram()
        histlength  = sum(histogram)
        probability = [float(h) / histlength for h in histogram]
        return -sum([p * math.log(p, 2) for p in probability if p != 0])



class Images3:

    def __init__(self, image1, image2, threshold=8):
        self.image1 = image1
        if type(image1) == str:
            self.image1 = sys.modules["Image"].open(self.image1)
        self.image2 = image2
        if type(image2) == str:
            self.image2 = sys.modules["Image"].open(image2)
        self.threshold = threshold

    def DoComparison(self, image1=None, image2=None):
        if not image1: image1 = self.image1
        if not image2: image2 = self.image2
        image = image1
        monoimage1 = image1.convert("P", palette=sys.modules["Image"].ADAPTIVE, colors=2)
        monoimage2 = image2.convert("P", palette=sys.modules["Image"].ADAPTIVE, colors=2)
        imgdata1   = monoimage1.getdata()
        imgdata2   = monoimage2.getdata()

        changed = 0
        i = 0
        acc = 3

        while i < DEFAULT_DEVICE_WIDTH * DEFAULT_DEVICE_HEIGHT:
            now  = imgdata1[i]
            prev = imgdata2[i]
            if now != prev:
                x = (i % DEFAULT_DEVICE_WIDTH)
                y = (i / DEFAULT_DEVICE_HEIGHT)
                try:
                    #if self.view == "normal":
                    image.putpixel((x,y), (0,0,256))
                    #else:
                    #    monoimage.putpixel((x,y), (0,0,256))
                except:
                    pass
                changed += 1
            i += 1
        percchange = float(changed) / float(DEFAULT_DEVICE_WIDTH * DEFAULT_DEVICE_HEIGHT)
        return percchange


if __name__ == "__main__":
    # image1 & image2 MUST be legit paths!
    image1 = "C:\\Path\\To\\Your\\First\\Image.jpg"
    image2 = "C:\\Path\\To\\Your\\Second\\Image.jpg"

    print "Images Result:"
    print Images(image1, image2).DoComparison()

    print "\nImages2 Result:"
    print Images2(image1, image2).DoComparison()

    print "\nImages3 Result:"
    print Images3(image1, image2).DoComparison()

Answer 1

这可能是个简单的方法，但这是一个很好的起点。我相信你会受到相机噪声的影响，而且你可能想要区分光线变化和图像构图变化。不过，这里是我想到的：

你可以使用PIL的ImageChops来高效地计算两张图片之间的差异。然后，你可以计算这个差异的熵，从而得到一个单一的阈值。

看起来这个方法是有效的：

from PIL import Image, ImageChops
import math

def image_entropy(img):
    """calculate the entropy of an image"""
    # this could be made more efficient using numpy
    histogram = img.histogram()
    histogram_length = sum(histogram)
    samples_probability = [float(h) / histogram_length for h in histogram]
    return -sum([p * math.log(p, 2) for p in samples_probability if p != 0])

# testing..

img1 = Image.open('SnowCam_main1.jpg')
img2 = Image.open('SnowCam_main2.jpg')
img3 = Image.open('SnowCam_main3.jpg')

# No Difference
img = ImageChops.difference(img1,img1)
img.save('test_diff1.png')
print image_entropy(img) # 1.58496250072

# Small Difference
img = ImageChops.difference(img1,img2)
img.save('test_diff2.png') 
print image_entropy(img) # 5.76452986917

# Large Difference
img = ImageChops.difference(img1,img3)
img.save('test_diff3.png')
print image_entropy(img) # 8.15698432026

我认为这是一个更好的图像熵算法，因为它在颜色空间中进行三维分箱，而不是为每个通道创建单独的直方图。

编辑 - 这个函数在2012年4月6日进行了修改

import numpy as np
def image_entropy(img):
    w,h = img.size
    a = np.array(img.convert('RGB')).reshape((w*h,3))
    h,e = np.histogramdd(a, bins=(16,)*3, range=((0,256),)*3)
    prob = h/np.sum(h) # normalize
    prob = prob[prob>0] # remove zeros
    return -np.sum(prob*np.log2(prob))

这些是我的测试图片：

图片1

图片2

图片3