Pedestrian-Motorcycle Binary Classification Using Data Augmentation and Convolutional Neural Networks

Abstract

One common problem in vehicle and pedestrian detection algorithms is the mis-classification of motorcycle riders as pedestrians. This paper focused on a binary classification technique using convolutional neural networks for pedestrian and motorcycle riders in different road context locations. The study also includes a data augmentation technique to address the un-balanced number of training images for a machine learning algorithm. This problem in un-balanced data sets usually cause a prediction bias, in which the prediction for a learned data set usually favors the class with more image representations. Using four data sets with differing road context (DS0, DS3-1, DS4-3, and DS4-3), the binary classification between pedestrian and motorcycle riders achieved good results. In DS0, training accuracy is 96.96% while validation accuracy is 81.52%. In DS3-1, training accuracy is 93.17% while validation accuracy is 86.58%. In DS4-1, training accuracy is 94.42% while validation accuracy is 97.00%. In DS4-3, training accuracy is 95.94% while validation accuracy is 88.59%.