siatwangmin / 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
• a color transform and append binned color features to my HOG feature vector.　features　normalization and randomization of training and testing dataset are conducted
• a sliding-window technique is adopted to propose candidate regions for car, and the regions are input into a trained classifier
• a pipeline works on a video stream and a heat map is created for recurring detections frame by frame to reject outliers and follow detected vehicles.
• Estimate a bounding box for vehicles detected.

• CarND-Vehicle-Dection.ipynb is the project solution , all the code contains in it.
• svc_pickle.p is the model and feature parameters
• README.md it the writeup file of the project
• test_result.mp4 is the performance on the test_video.mp4
• result.mp4 is the performance on the project_video.mp4

I started by reading in all the vehicle and non-vehicle images. there are 8792 cars and 8968 non-cars for training and testing, Here is some examples of vehicle and non-vehicle classes:

Here is an example using the RGB color space and HOG parameters of orientations=8, pixels_per_cell=(8, 8) and cells_per_block=(2, 2):

The 6th cell of CarND-Vehicle-Dection.ipynb show how I settel on the final choice of HOG parameters, I get a smaller data set of car and notcar, small_car number is 1125 and small_notcar is 1125, to investigate the best HOG parameters.

Only to consider the HOG features, experiments are conducted and the result is shown as below:

we can find that RGB color space is not good features for hog to detect vehicles, Time cost increases with the orientation numbers, Orientation num of 12 performorms good at aspect of accuracy and time costs.

but we still cannot find which Color space is the best. So I Run the HOG+SVM struction for the whole 8792 cars and 8968 non-cars data set

and we find that HOG features perform the best in YUV Color space on all vehicles and not vehicels data

so next we want to found the best Parameters and we run the HOG+SVM classifier on the smaller data set to get the intuition.

We can find from tables above that 8 pixels per cell is the best perfmance

We can find from tables above that 2 cells per block will be chosen

as the table shows below,the best spaitial size and hist_bins is (32, 32) and (32)　repectly.

So the final parameter is that

color_space = 'YUV' # Can be RGB, HSV, LUV, HLS, YUV, YCrCb
orient = 12  # HOG orientations
pix_per_cell = 8 # HOG pixels per cell
cell_per_block = 2 # HOG cells per block
hog_channel = "ALL" # Can be 0, 1, 2, or "ALL"
spatial_size = (32, 32) # Spatial binning dimensions
hist_bins = 32    # Number of histogram bins
spatial_feat = True # Spatial features on or off
hist_feat = False # Histogram features on or off
hog_feat = True # HOG features on or off

Since HOG + SVM Frame has achived a hug success in pedestrain detection, so in "Train Classifier" section of CarND-Vehicle-Dection.ipynb, I trained a linear SVM with the default classifier parameters and using HOG features with color spatial hist features and achives the accuracy of 99.01%

In " Sliding Window Search" section, I write a method find_cars based on the udacity lesson fuction. find_cars is a Hog Sub-sampling Window Search method, it's a combination of HOG feature extraction and a sliding window search to make the process more effcient and time-saving.

the HOG features are extracted for Region of Interest only once, and then features are subsampled according to the size of the window and then fed to the classifier.

Since the car looks smaller when far away and look larger when it's nearby, we should different scale in the different region. The region is show below:

The image below shows the 1st row of the table's boxes region:

The image below shows the 6st row of the table's boxes region:

Ultimately I searched on two scales using YUV 3-channel HOG features plus spatially binned color in the feature vector, which provided a nice result. The image below shows the rectangles returned by find_cars drawn onto test image in the final implementation. 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 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:

It takes lots of time to tune the parameter of HOG features, Finnaly I set a small data sets to find the intuition. I think it's is a good choice to do experiements on smaller data sets.

Particle filter or kalman filter can be added to the future work, Since there are correlations between frame and frame, this will make the work more robust.

SSD or Faster RNN can be a good direction to this project.