Object detection¶

Tracking (following something over time) typically starts with object detection (localizing the thing in a given image)

Background subtraction

What is “background”?

Thresholding for binarized images

Connected component labeling

Image: Chao..Perán Sensors 2015

Connected component labeling¶

https://en.wikipedia.org/wiki/Connected-component_labeling

248px-Two-pass_connected_component_labeling.svg.png Image: https://commons.wikimedia.org/wiki/File:Two-pass_connected_component_labeling.svg

Blur and erosion¶

Operations to help thresholding¶

Before thresholding: Gaussian blur

Image: Ronneberger..Joffe Chromosome Research 2008

Note that this is an example of kernel based image processing). This is also called "convolution".

Image: https://de.m.wikipedia.org/wiki/Datei:Halftone,_Gaussian_Blur.jpg

After thresholding: Erosion, dilation

See wikipedia on Erosion). Note the similarities to the kernel operations.

Image: Cereser, PhD thesis 2016

In [1]:

import numpy as np
import matplotlib.pyplot as plt
from scipy import ndimage
import imageio # "pip install 'imageio[ffmpeg]'"

In [8]:

# Here we define a helper function which we call below to make sure
# an image is an "unsigned 8 bit integer". This way, we know they
# take only a single byte per pixel and have a value from 0..255.

def ensure_dtype_uint8(arr):
    return arr.astype(np.uint8)

In [9]:

fname = 'short-movie20170810_182130.mp4'
reader = imageio.get_reader(fname)
for frame in reader:
    frame0 = frame[:,:,1] # take only green channel
    break

In [10]:

frame0.ndim

Out[10]:

In [11]:

frame0.shape

Out[11]:

(1024, 1280)

In [12]:

plt.imshow(frame0,cmap='gray')
plt.colorbar();

In [13]:

plt.imshow(frame0)
plt.colorbar();

In [14]:

# get all frames into big 3D array
(height,width) = frame0.shape
all_frames = []
reader.set_image_index(0) # return to start of file
for frame in reader:
    all_frames.append( frame[:,:,1] )
all_frames = np.array(all_frames)
print(all_frames.shape)
n_frames = len(all_frames)

(102, 1024, 1280)

In [15]:

mean_frame = np.mean(all_frames, axis=0)
median_frame = np.median(all_frames, axis=0)

In [16]:

plt.imshow(mean_frame, cmap="jet")
plt.colorbar();

In [17]:

plt.imshow(median_frame, cmap="jet")
plt.colorbar();

In [18]:

start_row, stop_row = (750,950)
start_col, stop_col = (400,800)

In [19]:

frame = all_frames[0,:,:]

plt.imshow(frame[start_row:stop_row,start_col:stop_col], cmap='jet')
plt.colorbar();

In [20]:

plt.imshow(median_frame[start_row:stop_row,start_col:stop_col], cmap='jet')
plt.colorbar();

In [21]:

plt.imshow(mean_frame[start_row:stop_row,start_col:stop_col], cmap='jet')
plt.colorbar();

In [22]:

frame_absdiff = abs(frame - median_frame)

plt.imshow(frame_absdiff[start_row:stop_row,start_col:stop_col], cmap='jet')
plt.colorbar();

In [23]:

threshold = 70

In [24]:

tmp = frame_absdiff>threshold
print(tmp.dtype)
ensure_dtype_uint8(frame_absdiff>threshold).dtype

bool

Out[24]:

dtype('uint8')

In [25]:

binarized = ensure_dtype_uint8(frame_absdiff>threshold)
plt.imshow(binarized[start_row:stop_row,start_col:stop_col], cmap='jet')
plt.colorbar();

In [26]:

blurred = ndimage.gaussian_filter(frame, sigma=6.0)
plt.imshow(blurred[start_row:stop_row,start_col:stop_col], cmap='jet')
plt.colorbar();

In [27]:

blurred_absdiff = abs(blurred - median_frame)
blur_threshold = 40
binarized_blurred = ensure_dtype_uint8(blurred_absdiff>blur_threshold)
plt.imshow(binarized_blurred[start_row:stop_row,start_col:stop_col], cmap='jet', interpolation='nearest')
plt.colorbar();

In [28]:

eroded = binarized_blurred
for i in range(4):
    eroded = ndimage.binary_erosion(eroded)
plt.imshow(eroded.astype(np.uint8)[start_row:stop_row,start_col:stop_col], cmap='gray')
plt.colorbar();

1.1 MiB Raw Blame History

Object detection¶

Connected component labeling¶

Blur and erosion¶

Operations to help thresholding¶

1.1 MiB

Raw Blame History