A new pedestrian detection method based on combined HOG and LSS featuresby Shihong Yao, Shaoming Pan, Tao Wang, Chunhou Zheng, Weiming Shen, Yanwen Chong

Neurocomputing

About

Year
2015
DOI
10.1016/j.neucom.2014.08.080
Subject
Artificial Intelligence / Computer Science Applications / Cognitive Neuroscience

Similar

MutualCascade method for pedestrian detection

Authors:
Hulin Kuang, Yanwen Chong, Qingquan Li, Chunhou Zheng
2014

HOG based fast human detection

Authors:
M. Kachouane, S. Sahki, M. Lakrouf, N. Ouadah
2012

Pedestrian Detection Based on Active Basis

Authors:
Sheng Ke Wang, Lan Xu, Xiaowei Xu
2012

A Discriminative Method for Pedestrians Detection on Real-Time Video

Authors:
Xue Yuan Chen, Xia Fu Lv, Jie Liu
2014

Text

A new pedestrian detection method based on combined

HOG and LSS features

Shihong Yao a, Shaoming Pan a, Tao Wang a, Chunhou Zheng b,

Weiming Shen a, Yanwen Chong a,n a State Key Laboratory for Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, 129 Luoyu Road, Wuhan 430079, China b College of Electrical Engineering and Automation, Anhui University, Hefei 230039, Anhui, China a r t i c l e i n f o

Article history:

Received 15 November 2013

Received in revised form 15 July 2014

Accepted 6 August 2014

Communicated by D.-S. Huang

Keywords:

Pedestrian detection

Feature combination

SVM

Adaboost a b s t r a c t

Pedestrian detection is a critical issue in computer vision, with several feature descriptors can be adopted. Since the ability of various kinds of feature descriptor is different in pedestrian detection and there is no basis in feature selection, we analyze the commonly used features in theory and compare them in experiments. It is desired to find a new feature with the strongest description ability from their pair-wise combinations. In experiments, INRIA database and Daimler database are adopted as the training and testing set. By theoretic analysis, we find the HOG–LSS combined feature have more comprehensive description ability. At first, Adaboost is regarded as classifier and the experimental results show that the description ability of the new combination features is improved on the basis of the single feature and HOG–LSS combined feature has the strongest description ability. For further verifying this conclusion, SVM classifier is used in the experiment. The detection performance is evaluated by miss rate, the false positives per window, and the false positives per image. The results of these indicators further prove that description ability of HOG–LSS feature is better than other combination of these features. & 2014 Published by Elsevier B.V. 1. Introduction

With the development of intelligent transportation, video surveillance and intelligent analysis, detecting pedestrians in image has become a vital research direction. Pedestrian features have more variable appearance and the wide range of poses comparing with other object features. That’s the reasonwhy the pedestrian detection is a challenging task [1]. The most critical part in pedestrian detection is how to extract effective pedestrian features on which the overall performance of detection depends.

There is much ongoing research in exploring a novel pedestrian feature for pedestrian detection. These features include Haar features [2], Scale invariant feature transform (SIFT) [3], speeded-up robust feature (SURF) [4], histogram of oriented gradient (HOG) [5], local binary pattern (LBP) features [6], local self-similarity (LSS) features [7], covariance features [8], multi-scale orientation (MSO) features [9], shapelet features [10] etc. The rigid deformation, perspective deformation of the image can be solved by SIFT features. But their high computational complexity hinders their further development. This limitation can be solved by SURF. In order to seek the lower computational complexity of features, Haar features were born. The computational procedure of Haar feature is very simple but it is easily affected by complex background. HOG features are robust for changes of clothes color, human body shape and height. LBP features are originally used for text classification and LSS features are obtained by calculating local self-similarity descriptors in dense and commonly used in image matching. MSO features, derived from Haar-like and

HOG features, are extracted on square image blocks with various sizes (called units) and contain coarse and fine features. The selected features are generally evaluated with linear support vector machine (SVM) [11] or Adaboost [12]. With the commercialization and productization of pedestrian detection, the requirement of detection performance is increasing. A single feature no longer meets the needs of the pedestrian detection precision.

Now, most scholars not only focus on pedestrian feature selection, but also devote themselves to studying the features improvement or the features combination [13,14]. Tomokia et al. [15] propose a method for extracting feature descriptors consisting of co-occurrence concept based on HOG (CoHOG). CoHOG is histograms of which a building block is any pair of gradient orientations.

The pair of gradient orientations has more information than single

Contents lists available at ScienceDirect journal homepage: www.elsevier.com/locate/neucom

Neurocomputing http://dx.doi.org/10.1016/j.neucom.2014.08.080 0925-2312/& 2014 Published by Elsevier B.V. n Corresponding author. Tel.: þ86 27 68771835; fax: þ86 27 68778229.

E-mail addresses: yao_shi_hong@whu.edu.cn (S. Yao), pansm@whu.edu.cn (S. Pan), wangtao.mac@whu.edu.cn (T. Wang), ywchong@whu.edu.cn (Y. Chong).

Please cite this article as: S. Yao, et al., A new pedestrian detection method based on combined HOG and LSS features, Neurocomputing (2014), http://dx.doi.org/10.1016/j.neucom.2014.08.080i

Neurocomputing ∎ (∎∎∎∎) ∎∎∎–∎∎∎ one. Including co-occurrence with various positional offsets, the feature descriptors can express complex shapes of objects with local and global distributions of gradient orientations. CoHOG can express more various shapes than HOG, which uses single gradient orientation. Because the rectangle detection window cannot handle rotation transformation and the pedestrian must be in upright pose due to the limitation in geometrical variation, Panachit et al. [16] suggest to use square-shaped window as the detection window, which can contain more variations of pedestrian. Original LBP descriptor does not suit the pedestrian detecting problem well due to its high complexity and lacking of semantic consistency, so

Mu et al. [17] propose two variants of LBP: Semantic-LBP and

Fourier-LBP. Liu et al. [18] propose two new texture features called local self-similarities (LSS, C) and fast local self-similarities (FLSS, C) based on Cartesian location grid, which achieve more robust geometric translations invariance and less computing time. These researches show that a single feature has some limitation and deficiency, and seeking feature combination has become a research focus in pedestrian detection. Wang et al. [19] combine the trilinear interpolated HOG with LBP as the feature set. This novel pedestrian detection approach is capable of handling partial occlusion and outperforms other state-of-the-art detectors on the INRIA dataset.