Vehicle Detection Project

The goals / steps of this project are the following:

• Perform a Histogram of Oriented Gradients (HOG) feature extraction on a labeled training set of images and train a classifier Linear SVM classifier
• Apply a color transform and append binned color features, as well as histograms of color, to your HOG feature vector.
• Implement a sliding-window technique and use your trained classifier to search for vehicles in images.
• Run your pipeline on a video stream (start with the test_video.mp4 and later implement on full project_video.mp4) and create a heat map of recurring detections frame by frame to reject outliers and follow detected vehicles.
• Estimate a bounding box for vehicles detected.

The whole project code is contained in the ipython notebook. The first half of the notebook covers training the classifier on the udacity image set (~8000 vehicle/non-vehicle images) The second half of the notebook has code for the video pipeline.

I used the udacity dataset which contains ~8000 vehicle and ~8000 non-vehicle images. I used the complete dataset.

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

I then explored different color spaces and different skimage.hog() parameters (orientations, pixels_per_cell, and cells_per_block).

Here is an example using the LUV color space and HOG parameters of orientations=8, pixels_per_cell=(8, 8) and cells_per_block=(2, 2) based on sample image shown above. The get_hog_features function handles this:

I also used features like color histogram and spatial binning Here is the visualization of the color historam for the 3 channels (using 32 bins):

I tried various combinations of parameters so as the get the best accuracy on the test set. My final parameter set was:

color_space = 'LUV' # Can be RGB, HSV, LUV, HLS, YUV, YCrCb

orient = 8 # HOG orientations

pix_per_cell = 8 # HOG pixels per cell

cell_per_block = 2 # HOG cells per block

hog_channel = 0 # Can be 0, 1, 2, or "ALL"

spatial_size = (16, 16) # Spatial binning dimensions

hist_bins = 32 # Number of histogram bins

spatial_feat = True # Spatial features on or off

hist_feat = True # Histogram features on or off

hog_feat = True # HOG features on or off

I trained a linear SVM using scikits linerSVC. I split up the dataset and used 20% as test dataset and remaining for training. My classifier achived an accuracy of ~0.9794

For sliding window search, I basically had three zones with differing window sizes. The window size in the farthest zone was the smallest. The search ignored all of the top half of the image (y coordinate less than 350). Three sets of windows were created - windows_far, windows_near, windows_mid. These are visualized on the test image below:

The below image shows the detections i got for the above sample image:

After generating the heatmap based on the detections, the following bounding boxes were visualized:

Here's a link to my video result

From detections I created a heatmap and then thresholded that map to identify vehicle positions. I then used scipy.ndimage.measurements.label() to identify individual blobs in the heatmap. I then assumed each blob corresponded to a vehicle.

In order to reduce the "jumpiness" of the boxes, I maintained a history of previous heatmaps. While processing every frame, I summed the history and then took a weighted average of the current heatmap and the history

The largest amount of time spent was on tuning the heatmap thresholds and incorporating historic frame information into current frame calculations (with appropriate weights) in order to make the bounding boxes smoother.

My classifer seems to work better on the darker car. This results in some sections of the video not detecting the white car. Part of the problem could be that udacity dataset may have less number of images for white cars. Although my classifier has accuracy of 0.979 , in some sampe images, number of detections for white car were less compared to the other car.