OpenCV Dashcam Car Detection
This repository contains a Haar classifier trained specifically to recognize the rear ends of cars, along with the entire data set used to train it.
Classifier Training Data
The classified was trained using stills from dashcam video, located in the positives and negatives directories. The stills in the negatives directory are from video footage containing no other cars on the road. There are currently around 450 negative images and 469 positive images of vehicles.
All of the positives were annotated with the bounds of the rear ends of cars in the frame
using the opencv_annotation tool (stored in annotations.txt).
A .vec file of samples was then created (24x24 sample size) using the opencv_createsamples
tool.
The Haar cascade was then trained to an acceptanceRatioBreakValue of 1.0e-5.
Retraining the Classifier
If you would like to train your own classifier using this data set, you can follow the steps
outlined in training_steps.sh. I would not recommend simply running that script, instead run
one command at a time.
You will need to delete the data in the cascade_dir first and change the numPos and numNeg input
to opencv_traincascade if you add new samples.
Example Python Code
If you have OpenCV installed you can run the Python sample code using python ./detect.py <path_to_video.mp4>