Journal of Intelligent & Fuzzy Systems 35 (2018) 4807–4820 DOI:10.3233/JIFS-18491 IOS Press 4807 On comprehensive analysis of learning algorithms on pedestrian detection using shape features Igi Ardiyanto ∗ , Teguh Bharata Adji and Dika Akilla Asmaraman Department of Electrical Engineering and Information Technology, Universitas Gadjah Mada, Yogyakarta, Indonesia Abstract. Despite the surge of deep learning, deploying the deep learning-based pedestrian detection into the real system faces hurdles, mainly due to the huge resource usages. The classical feature-based detection system still becomes feasible option. There have been many efforts to improve the performance of pedestrian detection system. Among many feature set, Histogram of Oriented Gradient seems to be very effective for person detection. In this research, various machine learning algorithms are investigated for person detection. Different machine learning algorithms are evaluated to obtain the optimal accuracy and speed of the system. Keywords: Pedestrian detection, machine learning, Histogram of Oriented Gradient, shape features 1. Introduction Person detection system has been continuously developed to get better performance. The system plays an important role in the latest technological developments, such as in Advanced Driving Assistant System. The success rate of this system is determined by its working accuracy and speed. Machine learning and feature extaction are the main factor behind the performance of the system. Many methods have been applied, both using conventional and modern meth- ods to achieve the best performance of the system. The conventional detection method using supervised machine learning and sliding window is preferred to build a low cost system. For many years, researchers have looked new ways to improve the performance of person detection sys- tem, both on the feature descriptor and on the machine ∗ Corresponding author. Igi Ardiyanto, Department of Electri- cal Engineering and Information Technology, Universitas Gadjah Mada, Jl. Grafika No. 2, Yogyakarta, Indonesia. E-mail: igi@ugm. ac.id. learning. For instance, Dalal et al. [1] had intro- duced new feature set known as HOG (Histogram of Oriented Gradient) that outperforms existing feature sets, such as Haar wavelet, Shape Context and PCA- SIFT (Principal Component Analysis-Scale Invariant Feature Transform) for linear SVM (Support Vector Machine) based person detection. The HOG feature set then becomes very popular because of its superior performance and is widely used for many detection problems, including person detection. Beside HOG, many feature sets, i.e. Haar wavelet and SIFT also have been used for person detection. For instance, [2, 3] use SVM classifier for person detection based on Haar wavelet. In [4], AdaBoost algorithm is used to learn a cascade of classifier based on spatial temporal Haar wavelet. In [5], the authors use recognition-by-component approach based on Haar wavelets and SVM training. The authors in [6] use two SVM classifiers and gradient magnitude fea- tures to detect person in front, rear, and side views. In [7], recognition-by-component approach is also used for detection based on SIFT features and AdaBoost 1064-1246/18/$35.00 © 2018 – IOS Press and the authors. All rights reserved