Pedestrian Detection of Road Scenes using Depth and Intensity Features Sisay Shimelis Gurmu School of Computer Science and Engineering, Kyungpook National University 80 Daehak-ro, Bukgu, Daegu, Republic of Korea sisay@vr.knu.ac.kr Min Woo Park School of Computer Science and Engineering, Kyungpook National University 80 Daehak-ro, Bukgu, Daegu, Republic of Korea mwpark@vr.knu.ac.kr Soon Ki Jung School of Computer Science and Engineering, Kyungpook National University 80 Daehak-ro, Bukgu, Daegu, Republic of Korea skjung@knu.ac.kr ABSTRACT In this paper, we present pedestrian detection method using fusion of intensity and depth features. Complementary fusion of these features significantly boosts the detection performance. Histogram of Oriented gradient (HOG) is applied for feature extraction in both intensity and depth images and trained by linear SVM. Our approach has an advantage over the conventional intensity image based methods, since depth features are robust against illumination, complex background and human pose variations. The experimental result shows that our proposed method has better detection performance. Categories and Subject Descriptors I.4.8 [Image Processing and Computer Vision]: Scene Analysis-Object recognition, Photometry General Terms System, Algorithm Keywords Pedestrian, Detection, Histogram of Oriented gradient (HOG), Depth Feature and Intensity Feature. 1. INTRODUCTION Pedestrians are the most susceptible members in urban traffic, thus the first step to protect pedestrians is to reliably detect them. Pedestrian protection has got much consideration in recent years and significant researches have been conducted by several researchers. Detecting pedestrians is applicable in a variety of important applications, especially in the perspective of intelligent transportation systems (ITSs) and intelligent vehicles (IVs) since it plays an important role for the pedestrian safety. Human detection is one of a challenging area in computer vision research filed due to enormous variation of human appearance and pose, illumination and complex background. The pedestrian detection under different circumstances introduces a multitude of complicating factors that make it one of the most challenging problems in computer vision. These complicating factors have to be accredited and addressed by computer vision systems if robust pedestrian detection is to become possible in real world scenarios. This work strives to overcome some of these challenges as much as possible by using stereo depth data especially depth information which has an advantage to avoid illumination and clattered background. We used our road scene stereo dataset to test our method and Daimler training dataset used to train the SVM [1]. We need stereo vision to get three-dimensional information from two images taken from slightly different viewpoints i.e. left and right cameras. From a computational standpoint, a stereo system must solve two problems. The first consists in determining which item in the left camera corresponds to which item in the right one (correspondence problem). The second problem is stereo reconstruction, if we solve the first problem; stereo reconstruction is straightforward by using a pure algebraic approach [2, 11]. Accordingly, most of the efforts have to be focused on finding the correct correspondences between image points from the left and right cameras. If the geometry of the stereo pair is known for each point the searching area in the other image is constrained to a single epipolar line. Most of the approaches to stereo vision assume that the epipolar lines run parallel to the image lines. This situation can be forced by means of stereo pair rectification. In this paper we used the stereo matching code of [2] to obtain depth map from our stereo road scene. The rest of this paper is organized as follows. In section 2, we introduce related works, in this section different human detection methods and some related works on object detection studied. Chapter 3 presents a technique to combine depth and intensity HOG features. Section 4 shows some experimental results, and finally Chapter 5 presents the conclusion of this paper. 2. RELATED WORKS Due to an enormous number of prospective applications, pedestrian detection has become one of an active research area in computer vision. As a result of this it has been a substantial amount of previous state of the art pedestrian detection techniques are proposed using various approaches. Techniques include the application of classic 2D computer vision techniques and the use Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. RACS’14, October 5–8, 2014, Baltimore, MD, USA. Copyright 2014 ACM 978-1-4503-3060-2/14/10 …$15.00.