Vehicle-Detection

Vehicle-Detection

Detect vehicles in a video feed.

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
• Optionally, you can also apply a color transform and append binned color features, as well as histograms of color, to your HOG feature vector.
• Note: for those first two steps don’t forget to normalize your features and randomize a selection for training and testing.
• 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.

Code files description:

• helper_functions.py has functions defined to reduce the code of other files.

• hog_test.py extracts hog, spatial, hist features & outputs the visualization of it.

Histogram of Oriented Gradients (HOG)

1. HOG features extraction:

single_img_features function in helper_fuctions.py provides the code for the feature extraction.

I started by exploring 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 are the output images with HOG parameters of ‘orient = 9’, ‘pix_per_cell = 8’, ‘cell_per_block = 2’, ‘hog_channel = 1’, ‘spatial_size = (4, 4)’, ‘hist_bins = 32’:

‘HSL’ Colorspace: ‘HSV’ Colorspace: ‘RGB’ Colorspace: ‘YCrCb’ Colorspace: ‘YUV’ Colorspace:

2. Why I settled on the final choice of HOG parameters:

I tried various combinations of parameters and tried plotting histogram features for many random images for both car and non-car caterogy. Finally, the chosen parameters gave distinctictive results and better accuracy than other.

For colorspace, YCrCb and YUV were giving good results. So, I proceeded with YUV at first.

3. How I trained a classifier using selected features:

I started with extracting 3000 random car & noncar images for spatial, color and hog features. Then, I normalized the features and splitted them to train and test features with ratio of 0.9:0.1. Then, I trained it with LinearSVC.

Sliding Window Search

1. How I implemented a sliding window search:

I decided to search the image for vehicle in Y range of [400, 720] and complete X range, because all of the vehicles appear in that region. Upper region in Y axis is filled with trees/skies. I started with window size of 64 for selected region. I tried to reduce false positives by adding heat threshold of 5.

2. How I decided what scales to search and how much to overlap windows:

Using parameters, scale = 1, heat_threshold = 5, window = 64, samples=8500, colorspace='YUV'. I got following image detection output:

As it was getting few false positives, I tried changing colorspace to 'YCrCb'. It had no false positives in test image detection. The output is:

3. Examples of test images to demonstrate how your pipeline is working:

Upon modifying parameters to: ystart = 400, ystops = [486, 502, 534, 646, 646], scales = [0.72, 1, 1.5, 2, 2.5], heat_threshold = 3, window = 64.

I get following output:

4. What did I do to optimize the performance of your classifier:

I used multiple windows to get more accurate detection as possible & added the threshold of 3 to remove the false detection.

scales = [0.72, 1, 1.5, 2, 2.5], heat_threshold = 3

Video Implementation

1. Link to my final video output:

Here’s a link to test video result & link to project video result

2. How I filtered false positives and implemented combination of overlapping bounding boxes:

I used the method as in image detection to filter the false positives with combining overlapping bounding windows in single frame with threshold of 6.

Discussion

Problems/issues I faced in the implementation of this project:

Getting detected one particular vehicle in image was a big problem. Here, the image with that vehicle:

When I tried to increase the threshold for heatmap, this vehicle was not getting detected. On decreasing the threshold the other image were getting 1-2 false positives.

So, I tested various scales & xstops for optimal detection. Finally, I was able to detect that particular vechicle.