Pedestrian Classification Using Self-Training Algorithm

Abstract

Object recognition is one of the key components in autonomous driving systems. However, many existing systems rely on large amounts of labeled training data which are expensive and difficult to obtain. To enhance the practicability of such system, it is essential to develop an object recognition algorithm that can learn from small number of labeled data in order to reduce the cost of data annotation. For this purpose, this paper presents an investigation on the feasibility of using a semi-supervised learning algorithm called self-training to reduce the number of labeled training data. To this aim, first, a self-training algorithm that uses artificial neural networks as a base classifier is prototyped to evaluate the performance of self-training algorithm alone. Next, self-training is implemented on top of a pedestrian recognition algorithm that adopts real Adaboost learning method with decision trees as weak classifiers. The feasibility of the proposed algorithm is confirmed by computational tests on the MNIST dataset of handwritten digits and the NICTA pedestrian dataset. The classification performances of traditional supervised learning and self-training algorithms trained under equivalent conditions are compared. The results show that self-training algorithm is feasible to reduce the number of labeled data by 78% while keeping the same recognition accuracy level of 95% on the condition that sufficient number of unlabeled data are given. In addition, the results indicate the possibility of replacing color images with grayscale images in object recognition to realize faster computation speed. Overall, it is confirmed that self-training can be applied to any existing classifiers as well as implemented on top of existing recognition algorithms.