异构特征与多视角姿态融合：行人检测的新突破

需积分: 9 127 浏览量更新于2024-08-26 收藏 1.78MB PDF 举报

本文主要探讨了一种针对视觉行人检测问题的创新方法，发表在《IEEE Transactions on Intelligent Transportation Systems》(VOL.16, NO.2, APRIL 2015)上。作者Wei Liu、Bing Yu、Chengwei Duan、Liying Chai、Huai Yuan和Hong Zhao针对行人检测中面临的挑战，如行人外观多样性、光照变化和部分遮挡，提出了一种结合异构特征与多视角姿态部件集合的行人检测技术。首先，异构特征的融合是核心部分。传统上，行人检测常用的是方向梯度直方图（Histogram of Oriented Gradients, HOG）和局部二值模式（Local Binary Pattern, LBP）这两种特征，它们分别对纹理和结构信息有良好的捕捉能力。然而，新方法在此基础上，设计了一种新颖的线性核函数，旨在更有效地整合这两种特性，增强行人描述符对光照条件和复杂背景的适应性。这种融合增强了特征表达的鲁棒性和区分度，有助于提高行人检测的准确性。其次，为了应对行人姿态变化和遮挡问题，文章提出了一个多视角-姿态部件集合（Multi-View-Pose Part Ensemble, MVPPE）检测器。这个系统利用了多个视角下不同身体部位的信息，通过集成学习的方式，使得模型能够更好地理解行人从不同角度和被遮挡时的视觉表现。通过这种方法，即便是在复杂的场景中，模型也能更准确地定位和识别行人。实验结果显示，该提出的特征组合策略显著提升了行人特征的描述能力，从而提高了行人检测的性能。在公共数据集上的测试表明，这种方法在面对各种挑战时表现出色，为视觉行人检测领域提供了一个有力的解决方案。这一研究成果对于提升智能交通系统的行人检测算法的鲁棒性和实用性具有重要意义，也为其他领域的目标检测任务提供了新的思路和借鉴。

LIU et al.: PEDESTRIAN-DETECTION BASED ON HETEROGENEOUS FEATURES AND ENSEMBLE OF MVP PARTS 815

Fig. 1. General architecture of the proposed pedestrian detector.

A. Overview of the Proposed Pedestrian-Detection Method

The architecture of our proposed method is presented in

Fig. 1. Due to high variability in pedestrian appearance, the

pedestrian is divided into several body parts, and each body

part is treated from different viewing angles and poses, re-

spectively. The details of division of parts, poses, and views

are given in Section III-C and D. For each view or pose of

a certain body part, an expert classiﬁer with heterogeneous

features (to be introduced in Section III-B) is trained. These

classiﬁers are assembled within a two-stage structure. The

ﬁrst stage ensembles different views and poses of each body

part, with view–pose ensemble (VPE) functions, and forms

an MVP ensemble classiﬁer for each body part. The second

stage combines all MVP body parts with a part ensemble

(PE) function. When an ROI is inputted to the detector, all

individual expert classiﬁers examine the ROI from their own

ﬁeld of expertise, i.e., a certain viewing angle or pose of

a certain body part. After collecting the opinions from the

experts, the VPE functions combine the classiﬁcation results

of body parts. Then, the PE function addresses the ﬁnal

decision result.

B. Pedestrian Feature Description Based on Combination of

Heterogeneous Features

One important step in the process of pedestrian detection is

to perform a thorough and distinctive feature description of the

pedestrian. The commonly used features include HOG, color

feature, LBP, Haar wavelet, and motion feature. One single

feature could describe only a single aspect of the pedestrian,

such as contour, color, local region, or texture, and it only

has limited description power. To perform a better description

of the pedestrian, some literatures propose to use the combi-

nation of more than one feature to enhance the description

power, such as HOG–LBP [18] and HOG–CSS features [15].

HOG–LBP features extract contour and texture information,

simultaneously, and are among the best performing (and most

popular) feature sets available [35], [36]. Nevertheless, the

simple concatenation of the two feature vectors, as in [18],

does not take the contributions of both individual features into

account, and the description ability of the features is not fully

exploited. Inspired by [37], in this paper, a new linear kernel

function is proposed to combine the two heterogeneous features

with complementary information, as

K(x

, x



k=0

(1 − β)x



k=0

βx

(1)

where K(x

, x

) represents the proposed kernel function; x

is the feature vector of sample i; x

=[x

]; x

rep-

resent the kth element of the feature vectors of HOG and

LBP, respectively; m and n are the dimensions of the feature

vectors of HOG and LBP. β is a combination coefﬁcient, which

determines the contribution of each feature, and β ∈ [0, 1].

With (1), the contour feature and the local region feature

are combined organically, with consideration of their respective

contributions. One could notice that the simple concatenation

approach proposed in [18] is a special case of (1), where

the contributions of two features are considered to be equal,

and β = 0.5. Compared to the method in [18], our approach

signiﬁcantly improves the description power of the feature

combination, without noticeable increase in computation cost.

In addition, compared to the RBF kernel function in [37],

our approach boosts less requirement of computation power.

Details are shown in Section IV-B.

In this paper, the extraction of HOG feature is the same

as [7]. LBP uses the same size of block (16 × 16) as HOG.

For each block, LBP generates a histogram with 58 uniform

patterns and 1 nonuniform pattern. The histograms of all blocks

are concatenated as the LBP feature vector of the input image.

As the size of the ROI is 48 × 96, the proposed heterogeneous

features describe the input image with a 5225-dimensional

feature vector.

C. Division of Body Parts

In order to handle possible partial occlusion, considering

both model complexity and detection accuracy, the pedestrian is

divided into three parts, i.e., UB, LB, and FB, which is the same

as the approach proposed in [9] and [25] (see Fig. 2). Every

part covers a ﬁxed percentage of the pedestrian. UB and LB

take 50% of the body, whereas FB covers 100% of the body.

The horizontally occluded pedestrian could be detected with

such division of the body. For example, the pedestrian with

an umbrella on his shoulder [UB occluded; see Fig. 3(a)] can

be properly detected by examining the nonoccluded LB. As for

the pedestrian with vertical occlusions [see Fig. 3(c) and (d)], as

long as the occlusion is less than 30% of the body, such kind of

occlusion could be handled with adequate vertically occluded

samples in the training data set.

剩余11页未读，继续阅读

weixin_38641366

粉丝: 4
资源: 893

异构特征与多视角姿态融合：行人检测的新突破

论文研究-基于异构节点的全视角强栅栏覆盖的研究.pdf

基于异构多核并行加速的嵌入式神经网络人脸识别方法.pdf

基于数据挖掘的异构网络多源目标数据融合跟踪方法研究.pdf

基于子空间学习方法的多视角聚类

基于异构图的图神经网络用于图像分类的完整代码示例

在异构CAD系统中，如何有效实现基于特征模型的几何元素标志转换以支持模型的集成和共享？

合成异构社交网络的方法

多源异构数据融合方法mcs-rf

如何实现异构CAD系统中的特征模型转换和几何元素的标志转换？

帮我写一篇基于多核异构处理器系统带宽调整策略专利

最新资源