The goals / steps of this project are the following:

Created numpy arrays with histogram values of all RGB colors and plotted it. Appended it to features set.

Resized the image and converted it to 1D for appending it to features set.

I started by reading in all the vehicle and non-vehicle images. Here is an example of one of each of the vehicle and non-vehicle classes:

I then explored different color spaces and different skimage.hog() parameters (orientations, pixels_per_cell, and cells_per_block). I grabbed random images from each of the two classes and displayed them to get a feel for what the skimage.hog() output looks like.

Here is an example using the YCrCb color space and HOG parameters of orientations=8, pixels_per_cell=(8, 8) and cells_per_block=(2, 2): Used skimage's hog() function to get hog features.

  features, hog_image = hog(img, orientations=orient, 
                            pixels_per_cell=(pix_per_cell, pix_per_cell),
                            cells_per_block=(cell_per_block, cell_per_block), 
                            transform_sqrt=True, 
                            visualise=vis, feature_vector=feature_vec)

I tried various combinations of color_space i.e, RGB, HSV, LUV, HLS, YUV, YCrCb, orientation i.e, 7, 9, 11 and cells per block as 2, 4. Observed the graphs and detection rate of vehicles. Finally settled with this set of values:

color_space = 'YCrCb' 
orient = 9  # HOG orientations
pix_per_cell = 8 # HOG pixels per cell
cell_per_block = 2 # HOG cells per block

I used Linear Support Vector Classifier for classifying the car and non-car from dataset. It gave over 99% accuracy.

Ultimate goal was to implement the algorithm on video frames. For this, I captured the image frames and created some windows/patches to slide over the image and find if the window contains car or not.

This was a pretty interesting work. I decided to use some random values. But it didn't work well. Some of the combinations are:

l =[]
w, h = 32, 32
for i in range(7):
    w += w*0.35
    w = int(w)
    h += h*0.25
    h = int(h)
    l.append((w,h))
print(l)

[(43, 40), (58, 50), (78, 62), (105, 77), (141, 96), (190, 120), (256, 150)]

Final Combination

w, h = 16, 16
l = []
for i in range(5):
    w += 24
    w = int(w)
    h += 24
    h = int(h)
    l.append((w,h))
print(l)

[(40, 40), (64, 64), (88, 88), (112, 112), (136, 136)]

I combined all the functions and coded heat map for removing redundant windows over single car.

Ultimately I searched on two scales using YCrCb 3-channel HOG features plus spatially binned color and histograms of color in the feature vector, which provided a nice result. Here are some example images:

Here's a link to my video result

I recorded the positions of positive detections in each frame of the video. From the positive detections I created a heatmap and then thresholded that map to identify vehicle positions. I constructed bounding boxes to cover the detected region by creating 1 single box from with center as centroid of all positive detections.

Here's an example result showing the heatmap from a series of frames of video, the result of scipy.ndimage.measurements.label() and the bounding boxes then overlaid on the last frame of video:

The results can still be improved. The detection of cars and objects in video are not smooth. It must be improved. I can try smoothening the boxes movement. I will delve bit more into it to help it attaining this kind of detection: Nvidia Driveworks