In this project, we will use write a software pipeline to detect vehicles in a video (start with the test_video.mp4 and later implement on full project_video.mp4) however the main product is to create a detailed writeup of the project.
- Histogram of Oriented Gradients (HOG)
- Feature extraction on a labeled training set of images
- Train a classifier Linear SVM classifier using selected HOG features
- Sliding Window Search (SWS)
- Implement a SWS technique
- Use trained classifier to search for vehicles in images.
- Video Implementation
- The SWS plus classifier has been used to search for and identify vehicles in the videos provided.
- Implement some kind of filter for false positives and some method for combining overlapping bounding boxes.
- Reflection
| Project Video |
|---|
 |
- Datasets provided by Udacity
- Car Images
- NonCar Images
- Example of random sample images from the vehicle and non-vehicle datasets.
HOG FINAL PARAMETERS
color_space = 'YCrCb'
orient = 9
pix_per_cell = 8
cell_per_block = 2
hog_channel = "ALL"
spatial_size = (16, 16) # Spatial binning dimensions
hist_bins = 16 # Number of histogram bins
spatial_feat = True # Spatial features on or off
hist_feat = True # Histogram features on or off
hog_feat = True # HOG features on or off
The get_hog_features function takes in an image and computes the Histogram of Oriented Gradient (HOG). Takes image as input and HOG parameters (orientations, pixels_per_cell, cells_per_block)
Sample Result of Different HOG parameters
| SVM Accuracy | Orientation | Pixels Per Cell | Feature Vector Length | Time |
|---|---|---|---|---|
| 98.59% | 11 | 16 | 2 | 4356 |
| 98.45% | 9 | 16 | 2 | 4140 |
| 97.38% | 11 | 16 | 2 | 4356 |
| 97.61% | 9 | 16 | 2 | 4140 |
| 98.62% | 9 | 8 | 2 | 8460 |
Color Space mainly used were RGB, YUV, YCrCb.
SVM Classifier with the default classifier parameters using HOG features as indicated above was able to achieve test accuracy of 98.62%. Features were scaled to zero mean and unit variance before training the classifier using sklearn's StandardScaler (slection for training/testing radomised using np's random).
Final Scales and window sizing
scales = [1, 1.5]
y_start_stops = [[380, 460], [380, 560]]
Different scales and windows found to be relevant as tried out in notebook
scales = [1, 1.5, 2, 2.5, 4]
y_start_stops = [[380, 460], [380, 560], [380, 620], [380, 680], [350, 700]]
After playing with sliding window the result was to use find_cars mthong from the lesson with bit of tweaking. cells_per_step = 1 used instead of overlap to define how many cells to step.
Performance was optimized by using
- heatmap
- threshold
- window sizing
-
Pipeline perform reasonably well on the entire project video.
-
Identifies vehicles most of the time.
-
Minimum or no flase positives infact it even identifies vehicles coming from other side at times.
-
Bit woobly and unstable bounding boxes at times.
- find_cars does add_heat for heatmap
- find_cars does apply_threshold for threashold
- find_cars which does cells_per_step to combining instead of overlapping for bounding boxes.
- process_image does scales and window sizing
- Balancing the accuracy of the classifier with execution speed was the key.
- If you do more on accuracy you start to drop boxes
- If you do more on speed you end having false positives.
- Probably do better in presence of shadows
- Boxes could be better transition form one frame to another.