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
- Implement a sliding-window technique and use your trained classifier to search for vehicles in images.
- Estimate a bounding box for vehicles detected.
- Run this pipeline on a video stream
I defined the f_hog
function to extract the hog feature (with the skimage.hog()
) for every image.
I set orientations = 11
, pixels_per_cell = 16
, and cells_per_block = 2
and used the yuv
I processed the vehicle/non-vehicle dataset(feature_extraction and standardization); then I split in train and cross-validation set using train_test_split
with stratification.
After this I trained a support vector machine with rbf
kernel
obtaining a 0.991
of accuracy with 3-fold cv.
I decided to search for different scales in the bottom part of the image: I tried, defining compute_windows
and collect_windows
, to compute the HOG feature for every windows, but it resulted to be a slow method; so in the find_subimages_boxes_features
I computed one time the features sub-sampling for every windows.
Here and example of the windows search space:
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 vehicles and I constructed bounding boxes to cover the area of each blob detected.
I defined my pipeline: extract the feature vector, compute the prediction for vehicle detection, define the bounding boxes using; finally I obtained:
For the video stream I implemented two method for false positive detection into the Detector()
class: the first is the thresholding on the heatmap; the second is that, if in the actual frame we identify a different number of cars than in the last frame, we don't identify a box and we wait for a new detection in the next frame: if in the next frame the classifier identifies again the same number of vehicles I consider a true positive detection
Here's a link to my video result
My approach is not robust and there are problem when the vehicle is on the right of the image; I see some problem also when a new vehicle enters in the image. In general I have some false negative. And I could improve my classifier.