Skip to content

Latest commit

 

History

History

Vehicle-Detection

README.md

Vehicle Detection

Udacity - Self-Driving Car NanoDegree

In this project, the goal is to write a software pipeline to detect vehicles in a video project_video.mp4. The project aims to detect the position of the vehicle in the image using conventional methods such as HOG(Histogram of Oriented Gradient) and SVM(Support Vector Machine).

pipeline

Usage

Dataset: Vehicles and Non-vehicles data

Udacity CarND provides datasets for vehicles and non-vehicles data to train classifier. It contains 8792 vehicle images and 8968 non-vehicle images. These datasets are comprised of images taken from the GTI vehicle image database, the KITTI vision benchmark suite, and examples extracted from the project video itself.

Train SVM Classifier

To train classifier, place images like:

    # for vehicle dataset:
    ./classifier_images/vehicles/xxx/yyy.png (folder/file name dosen't matter)

    # for non-vehicle dataset:
    ./classifier_images/non-vehicles/zzz/www.png

then follow Training_Classifier.ipynb.

Generate Video

To generate video, follow Generate_Video.ipynb.

Vehicle Detection Pipeline

You can find all the detail of vehicle detection pipeline on Vehicle_Detection_Pipeline.ipynb.

The algorithm is structured as follows:

  1. Extracting Features from images
    • Spartially Binned Color, Color Histogram, HOG
    • skimage.feature.hog() is used for extracting HOG features.
  2. Training Classifier
    • sklearn.svm.LinearSVC() function is used.
  3. Searching Using Sliding Window and Classifying Each Window
    • find_car() function is used for searching and classifying.
  4. Removing Multiple Detections and False Positives
    • apply_threshold() function and scipy.ndimage.measurements.label() function is used.

Result

It took 6 minutes to process 50 second video and generate the output video using moviepy. You can see full version of high resolution output video on Youtube: output_video

Input Result
input_video output_video

Discussion

Accuracy and Robustness

The accuracy of the classifier that distinguishes between vehicle and non-vehicle is about 98%. However, false possitives appear many times. I could reduce these false positives by using the previous frame's heatmap, but I could not remove them completely as shown in the image below. This false positive will be fatal when this detection algorithm is used for control the vehicle because The algorithm recognizes the empty road as obstacle.

Additional Readings

This project used a classic method of detecting vehicles in images. but state-of-the-art vehicle detection methods are proposed these days. In the future, I will apply the following techniques to this video.