CS 223B Project Presentation

1
CS 223B Project Presentation

• Detecting cars from an on-board moving vehicle
• Vamsi Vytla

2
Introduction

• Problem
• Identify cars from various backgrounds and
  lighting conditions
• Cars themselves can be of various colors, shapes,
  and sizes
• Results must tend to real-time, for any practical
  purposes
• Assumptions
• Camera located at the front of the car
• Offline detection (not real-time)

3
Approach

• Car detection is in a way similar to face
  detection
• Hence an approach similar to Viola and Jones
  paper
• Collect lots of samples of car rear-ends from
  Google - Street View 550 cropped car images,
  750 non-car images
• Reduce and normalize all samples to a much
  smaller scale
• Run an Adaboost like trainer that generates a
  strong classifier using various weak Haar
  features
• Evaluate the performance of the strong-classifier
• Higher-Level pruning of the detected windows
• Use the horizon line, temporal information from
  videos, size consistency

4
Results

• High number of false positives from the
  classifier A good thing, as other information
  can be applied to prune the hits
• The classifier tends to converge around 400-150,
  and 450-100 split of test data as it produces
  similar hit percentage and false-positive
  percentage (Notes below)
• Horizon line assumption further removes false
  positives

5
Samples of Input Images
Few Results
6
Conclusion Improvements

• A definite indication that cars are being
  detected robustly from images taken from the web,
  under various lighting conditions.
• Larger database of cars and non-cars should
  definitely improve results, and keep false
  positives lower
• Videos and further information of the camera will
  definitely aid in further pruning false positives
  and degenerate cases
• Detection with the classifier is very fast, and
  results are encouraging to be taken on videos and
  then real-time (Notes below)