sai19872000/vehicle-detection

Using machine learning to detect vechicles

Vehicle Detection

Prject Goals

The primary goals of this project are to:

1. Extract HOG features from test images for input into a classifier.

2. develop a classifier to detect presence of vehicle

3. Use a sliding window technique to scan through the image for the purpose of detecting a car

4. Combine the sliding window and classifier for decting and locating vehicles

5. Implement this in a video steam and use the previous frames reduce false positives

Data exploration

cell-4 in jupyter notebook: Firstly, I used the data provided far, left, middle_close, right point of view of cars additionally I have also used the images from KITTI. Here are the sample images randomly chooses from these sources

I used the get hog features cell-5 function using skimage library

from skimage.feature import hog

after extracting the hog features the image looks as follows

I also visualized the non car images that are used to train the classifier; after extracting the hog features we have the following

Exploring the color space and other parameters

I then started exploring the color space and other parameters such as: number of orientatons, pixels per cell, number fo cells perblock, hog channels and used a SVM classifier to make predictions. The case where there is resonable test accuracy is the classifier that I have used for making vehicle detection in video.

Here are some sample resutls cells(9 - 28):

Case	Colorspace	Number of orientations	Pixels Per Cell	Cells Per Block	HOG Channels	SVM Classifier Test Accuracy (%)
1	RGB	9	8	2	ALL	97.18
2	HSV	9	8	2	ALL	98.11
3	LUV	9	8	2	ALL	97.55
4	HLS	9	8	2	ALL	98.76
5	YCrCb	9	8	2	ALL	98.25
6	YUV	9	8	2	ALL	98.54
7	YUV	10	8	2	ALL	98.48
8	YUV	11	8	2	ALL	98.37
9	YUV	12	10	2	ALL	98.23
10	YUV	11	16	2	ALL	98.03

I have then used the

find_cars

function on cell 30 and the function "draw_boxes" function to draw boxes on the cars in the image. The figure after applying these function is as below

Sliding Window implementation

I have then used sliding window implementation used during the course. First I have used windows with several sizes different window sizes as shown below

after combining we have (cell 56):

Then I apply the "add_heat", apply_threshold, and "draw_labeled_bboxes" adapted from course to make heat maps of the vehicles.

Applying threshold we have

Changing the cmap to "gray"

Finally I then make the box that fits the vehicle

I then combined all the functions and a pipeline and then tested it on all the test images:

I used this pipeline and applied to both the test video and the project video and successfully identified the vehicles the link for the project video is ![here][https://github.com/sai19872000/vehicle-detection/blob/master/project_video_out.mp4]