ryannng / carnd-vehicle-detection Goto Github PK

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
• 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.

###Here I will consider the rubric points individually and describe how I addressed each point in my implementation.

###Histogram of Oriented Gradients (HOG)

The code for this step is contained in the fourth code cell of the IPython notebook .

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=9, pixels_per_cell=(8, 8) and cells_per_block=(2, 2):

Car

Codes for this section is called Experiment : Chooseing HOG Perimeter in ipython notebook.

I first extract a random sample consisting 1000 files from the car and notcar dataset for reducing experiment time.

Then I set several (3 or 4) candidate values for each parameter, such as [6, 9, 12] for orientation, then use for loop iterating each combination of parameters. It runs 4 (color space) * 3 (orientation) * 2(pix_per_cell) * 3 (cell_block) * 4 (color channels) = 288 times . After reviewing these test score, I got the insights that 'ALL' color channel is better than single channels, 'YCrCb'and 'HSV' perform better than other color space most of time.

Meantime, from the lesson's intructor's notes, the lecture shows some best practice:

I then narrow the search space, only provide 2 candidate value for each parameter. The final results shows that the combination of

# Best parameter
color_space = 'YCrCb'
orient = 18
pix_per_cell = 12
cell_block = 1
hog_channel = 'ALL'

performs best and got 0.99 score on the test dataset.

The code is entitled "SVC Model Traing" in the notebook.

I first gather all cars and notcars image from the dataset (at the begining of the notebook), then extract hog feature from these raw images. I use StandardScalor() from sklearn.preprocessing to normalize and reduce variance of the feature . After that I use train_test_split() with 0.2 ration of test set to split the whole dataset into training and testing data.

Then I import LinearSVC()model from sklearnto train on the HOG features, and use test data for validation. The final score is 0.9789

Finally, I store the svc model into pickle.

The code is in 7.Sliding Window. I first use slide_window()to collect window for one size. then define multi_scale_window() for generating multiple scale windows.

I decided to search the space in 4 scale: (64,64) for distant small object, (96, 96) for moderate distance object, large and extreme large for closest object.

the shape of cars on your right or left is a rectangle, so I set extreme large box ratio of width/height to 4:3.

####2. Show some examples of test images to demonstrate how your pipeline is working. What did you do to optimize the performance of your classifier?

Ultimately I searched on 4 scales using YCrCb 3-channel HOG features , which provided a nice result. Here are some example images:

Here's a link to my video result

(You may find the project_video_out.mp4 video in the forreport folder)

The code is in 8. Heat map

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.

There are still a lot of false positive boxes. I then take average of heat map of last 5 frames.

heat_list = deque(maxlen = 5)
zero = np.zeros_like(heat)
for heat in heat_list:
    zero += heat
heat_avg = zero / 5

I then used scipy.ndimage.measurements.label() to identify individual blobs in the heatmap. I then assumed each blob corresponded to a vehicle. I constructed bounding boxes to cover the area of each blob detected.

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:

###Discussion

I think that the biggest drawback of sliding window is that you have to pre-defined the window size and overlap rate, which may not generalize well. The pipeline in my implementation only process the currently frame thus sometimes fail to track the object all the way. And Heat map is a great way for eliminating false positives but may omit small or distant objects for lacking bounding boxes.

Here are 3 ideas for future improvements:

• Use selective search gererating window automatically rather than sliding window approach.
• Use NMS to reduce box overlapping
• Use HOG sub-sampling to reduce processing time