>>> 星瞳科技|OpenMV中国官方网站 <<<

>>> 星瞳科技店铺地址|OpenMV中国官方代理,全球九大代理商之一 <<<

>>> OpenMV3 Cam M7 官方高配版上市啦~戳我戳我戳我~ <<<

>>> 星瞳科技-OpenMV中文教程网 <<<

本例程利用 FAST/FREAK 算法进行特征提取,并且进行目标追踪。

注意:如果想要进行更加稳定的特征点检测,建议把目标物体的各个角度、各个大小的特征保存下来进行匹配。

# Freak Example
#
# This script shows off keypoint tracking by itself. Put an object in front of
# your OpenMV Cam without anything else in the image (i.e. camera should be
# facing a smooth wall) and the camera will learn the keypoints for an track
# whatever object is in the image. You can save keypoints to disk either via
# the OpenMV IDE or from in your script.
#
# Matching keypoints works by first extracting keypoints from an ROI. Once those
# are extracted then the OpenMV Cam compares the extracted keypoints against all
# the keypoints in an image. It tries to find the center matching point between
# the two sets of keypoints.
#
# Keep in mind that keypoint matching with just one training example isn't very
# robust. If you want professional quality results then stick with getting
# professionally generated Haar Cascades like the frontalface or eye cascade.
# That said, if you're in a very controlled enviroment then keypoint tracking
# allows your OpenMV Cam to learn objects on the fly.
#
# If... you want really good keypoint matching results we suggest you gather
# keypoints from all faces of an object and with multiple rotations and scales.
# Comparing against all theses sets of keypoints helps versus just one.
#
# NOTE: LOTS OF KEYPOINTS MAY CAUSE THE SYSTEM TO RUN OUT OF MEMORY!

import sensor, time, image

# Normalized keypoints are not rotation invariant...
NORMALIZED=False
#NORMALIZED=False,特征可旋转检测

# Keypoint extractor threshold, range from 0 to any number.
# This threshold is used when extracting keypoints, the lower
# the threshold the higher the number of keypoints extracted.
KEYPOINTS_THRESH=30
#设置特征点提取数目

# Keypoint-level threshold, range from 0 to 100.
# This threshold is used when matching two keypoint descriptors, it's the
# percentage of the distance between two descriptors to the max distance.
# In other words, the minimum matching percentage between 2 keypoints.
MATCHING_THRESH=80
#设置特征匹配阈值

# Reset sensor
sensor.reset()

# Sensor settings
sensor.set_contrast(1)
sensor.set_gainceiling(16)
sensor.set_framesize(sensor.QQVGA)
sensor.set_pixformat(sensor.GRAYSCALE)

# Skip a few frames to allow the sensor settle down
# Note: This takes more time when exec from the IDE.
for i in range(0, 30):
    img = sensor.snapshot()
    img.draw_string(0, 0, "Please wait...")
#等待一会儿,稍做准备

kpts1 = None
#kpts1保存目标物体的特征,可以从文件导入特征,kpts1 = image.load_descriptor(image.FREAK, "/desc.freak"),但是不建议这么做。
# Uncomment to load keypoints from file
#kpts1 = image.load_descriptor(image.FREAK, "/desc.freak")
clock = time.clock()

while (True):
    clock.tick()
    img = sensor.snapshot()
    kpts2 = img.find_keypoints(threshold=KEYPOINTS_THRESH, normalized=NORMALIZED)
    #提取目前摄像头的图像特征。find_keypoints(roi=Auto, threshold=32, normalized=False)
    #roi=(x,y,w,h)表示特征提取的范围,是左上顶点为(x,y),宽和高分别为w h 的矩形,默认是整个图像。
    #threshold为特征提取数目,阈值越大,提取的特征点数目越多,占用内存越大,有可能会超出内存。
    #normalized=False设置特征可旋转。

    if (kpts1==None):
    #如果是刚开始运行程序,提取最开始的图像作为目标物体特征,kpts1保存目标物体的特征
    #kpts2保存以后每帧图片的特征,方便与kpts目标特征进行比较。
        kpts1 = kpts2
        print(kpts1)
    elif kpts2:
    #如果不是刚开始运行程序,将当前图像与目标物体特征进行比较。
        c = image.match_descriptor(image.FREAK, kpts1, kpts2, threshold=MATCHING_THRESH)
        #match_descriptor(type, descritor0, descriptor1, threshold=60)
        #type表示图像特征,可以是image.LBP或者image.FREAK。descritor0和descriptor1
        #表示需要比较的两个图像的特征。threshold调整匹配阈值,阈值越大,匹配的相似度越高,
        #但是匹配成功的次数会变少。c[2]返回特征的相似度,如果相似度大于25%则认为匹配成功
        # C[2] contains the percentage of matching keypoints.
        # If more than 25% of the keypoints match, draw stuff.
        if (c[2]>25):
        #相似度大于25%,匹配成功
            img.draw_cross(c[0], c[1], size=15)
            #在匹配成功的目标中心画十字形,c[0] c[1]表示匹配的目标中心点,十字形大小为15
            img.draw_string(0, 10, "Match %d%%"%(c[2]))

    # Draw FPS
    img.draw_string(0, 0, "FPS:%.2f"%(clock.fps()))

运行程序效果: