What it does This notebook detects cars in images using hand-engineered computer vision. It extracts HOG features, color histograms, and spatially binned pixels, then feeds them into a linear SVM classifier. A sliding window scans across frames to localize vehicles, with bounding boxes drawn on the output.

The interesting bit This is pre-deep-learning computer vision in its pure form—no YOLO, no ResNet, just gradient histograms and a linear separator. The author deliberately subsamples training data (keeping only 1 in 5 images) to avoid overfitting, which is either prudent or a sign of how brittle feature engineering can be.

Key highlights

• Uses HOG, color histograms, and spatial binning concatenated into an 11,988-dimension feature vector
• Linear SVM trains in ~2.5 seconds on the reduced dataset
• Sliding window search with configurable scales and overlap
• Supports multiple color spaces (RGB, HSV, YUV, YCrCb, etc.)
• Includes a YouTube video demo of the pipeline running on road footage

Caveats

• Requires manually downloading GTI and KITTI datasets; nothing is bundled
• README is truncated mid-output, so final accuracy numbers and inference pipeline details are cut off
• Uses deprecated skimage HOG parameters that throw warnings

Verdict Worth a look if you’re learning classical CV or need to understand what YOLO replaced. Skip it if you want a production detector; this is coursework, not a framework.