Mathematics

2023-02-25 16:40| 来源: 网络整理| 查看: 265

Next Article in Journal A Gradient-Based Optimizer with a Crossover Operator for Distribution Static VAR Compensator (D-SVC) Sizing and Placement in Electrical Systems Previous Article in Journal A New Class of Generalized Probability-Weighted Moment Estimators for the Pareto Distribution Journals Active Journals Find a Journal Proceedings Series Topics Information For Authors For Reviewers For Editors For Librarians For Publishers For Societies For Conference Organizers Open Access Policy Institutional Open Access Program Special Issues Guidelines Editorial Process Research and Publication Ethics Article Processing Charges Awards Testimonials Author Services Initiatives Sciforum MDPI Books Preprints Scilit SciProfiles Encyclopedia JAMS Proceedings Series About Overview Contact Careers News Blog Sign In / Sign Up Notice clear Notice

You are accessing a machine-readable page. In order to be human-readable, please install an RSS reader.

Continue Cancel clear

All articles published by MDPI are made immediately available worldwide under an open access license. No special permission is required to reuse all or part of the article published by MDPI, including figures and tables. For articles published under an open access Creative Common CC BY license, any part of the article may be reused without permission provided that the original article is clearly cited. For more information, please refer to https://www.mdpi.com/openaccess.

Feature Papers represent the most advanced research with significant potential for high impact in the field. Feature Papers are submitted upon individual invitation or recommendation by the scientific editors and undergo peer review prior to publication.

The Feature Paper can be either an original research article, a substantial novel research study that often involves several techniques or approaches, or a comprehensive review paper with concise and precise updates on the latest progress in the field that systematically reviews the most exciting advances in scientific literature. This type of paper provides an outlook on future directions of research or possible applications.

Editor’s Choice articles are based on recommendations by the scientific editors of MDPI journals from around the world. Editors select a small number of articles recently published in the journal that they believe will be particularly interesting to readers, or important in the respective research area. The aim is to provide a snapshot of some of the most exciting work published in the various research areas of the journal.

Journals Active Journals Find a Journal Proceedings Series Topics Information For Authors For Reviewers For Editors For Librarians For Publishers For Societies For Conference Organizers Open Access Policy Institutional Open Access Program Special Issues Guidelines Editorial Process Research and Publication Ethics Article Processing Charges Awards Testimonials Author Services Initiatives Sciforum MDPI Books Preprints Scilit SciProfiles Encyclopedia JAMS Proceedings Series About Overview Contact Careers News Blog Sign In / Sign Up Submit Journals Mathematics Volume 11 Issue 5 10.3390/math11051075

Submit to this Journal Review for this Journal Edit a Special Issue ► ▼ Article Menu Article Menu Academic Editors

Pradeep Kumar Mallick

Arun Kumar Ray

Gyoo-Soo Chae Subscribe SciFeed Recommended Articles Related Info Link Google Scholar More by Authors Links on DOAJ Saleh, S. Abdulla Mohsen Kadarman, A. Halim Suandi, S. Azmin Ghaleb, S. A. A. Ghanem, W. A. H. M. Shuib, S. Hamad, Q. Shihab on Google Scholar Saleh, S. Abdulla Mohsen Kadarman, A. Halim Suandi, S. Azmin Ghaleb, S. A. A. Ghanem, W. A. H. M. Shuib, S. Hamad, Q. Shihab on PubMed Saleh, S. Abdulla Mohsen Kadarman, A. Halim Suandi, S. Azmin Ghaleb, S. A. A. Ghanem, W. A. H. M. Shuib, S. Hamad, Q. Shihab /ajax/scifeed/subscribe Article Views Citations - Table of Contents Altmetric share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up ... Endorse textsms ... Comment Need Help? Support

Find support for a specific problem in the support section of our website.

Get Support Feedback

Please let us know what you think of our products and services.

Give Feedback Information

Visit our dedicated information section to learn more about MDPI.

Get Information clear JSmol Viewer clear first_page settings Order Article Reprints Font Type: Arial Georgia Verdana Font Size: Aa Aa Aa Line Spacing:    Column Width:    Background: Open AccessArticle A Tracklet-before-Clustering Initialization Strategy Based on Hierarchical KLT Tracklet Association for Coherent Motion Filtering Enhancement by

Sami Abdulla Mohsen Saleh 1 ORCID

A. Halim Kadarman 2,*,

Shahrel Azmin Suandi 1,* ORCID

Sanaa A. A. Ghaleb 3,

Waheed A. H. M. Ghanem 4,

Solehuddin Shuib 5 and

Qusay Shihab Hamad 1,6 ORCID

1 Intelligent Biometric Group, School of Electrical and Electronic Engineering, Universiti Sains Malaysia, Nibong Tebal 14300, Pulau Pinang, Malaysia 2 School of Aerospace Engineering, Universiti Sains Malaysia, Nibong Tebal 14300, Pulau Pinang, Malaysia 3 Faculty of Computing and Informatics, Universiti Sultan Zainal Abidin, Kampung Gong Badak 21300, Terengganu, Malaysia 4 Faculty of Ocean Engineering Technology and Informatics, Universiti Malaysia Terengganu, Kuala Terengganu 21030, Terengganu, Malaysia 5 Faculty of Mechanical Engineering, Universiti Teknologi Mara, Shah Alam 40450, Selangor, Malaysia 6 Quality Assurance Department, University of Information Technology and Communications, Baghdad 10068, Iraq * Authors to whom correspondence should be addressed. Mathematics 2023, 11(5), 1075; https://doi.org/10.3390/math11051075 Received: 28 November 2022 / Revised: 7 February 2023 / Accepted: 8 February 2023 / Published: 21 February 2023 (This article belongs to the Special Issue Artificial Intelligence, Pattern Recognition and Data Learning with Applications in Engineering and Science) Download Download PDF Download PDF with Cover Download XML Browse Figures Versions Notes

Abstract: Coherent motions depict the individuals’ collective movements in widely existing moving crowds in physical, biological, and other systems. In recent years, similarity-based clustering algorithms, particularly the Coherent Filtering (CF) clustering approach, have accomplished wide-scale popularity and acceptance in the field of coherent motion detection. In this work, a tracklet-before-clustering initialization strategy is introduced to enhance coherent motion detection. Moreover, a Hierarchical Tracklet Association (HTA) algorithm is proposed to address the disconnected KLT tracklets problem of the input motion feature, thereby making proper trajectories repair to optimize the CF performance of the moving crowd clustering. The experimental results showed that the proposed method is effective and capable of extracting significant motion patterns taken from crowd scenes. Quantitative evaluation methods, such as Purity, Normalized Mutual Information Index (NMI), Rand Index (RI), and F-measure (Fm), were conducted on real-world data using a huge number of video clips. This work has established a key, initial step toward achieving rich pattern recognition. Keywords: crowd analysis; coherent motion detection; trajectory clustering; KLT tracklets MSC: 62H30; 65D19; 68T45; 37M10 1. IntroductionVideo surveillance plays a key role in the field of public safety management. In the case of gigantic crowd scenes, such as airports, shopping malls, stations, etc., using traditional monitoring methods cannot effectively supervise the behaviour of the crowd due to many influencing factors, including the large scale of the crowd, low resolution, serious occlusions, and complicated motion patterns [1,2,3]. Smart video surveillance systems, which are based on computer vision and image processing, can automatically complement various tasks. According to many survey papers [4], crowd analysis is subdivided into two research axes: crowd statistics and crowd behavior analysis. The purpose of crowd statistics is to estimate crowd density by the means of crowd-counting methods [5,6,7,8]. The purpose of crowd behavior analysis is to study the behavior of a crowd, such as a crowd motion detection and scene understanding [9,10,11,12,13,14,15,16,17], crowd event detection [18,19,20], and crowd anomaly detection [21,22,23,24]. In this regard, collective motion analysis has recently received considerable attention.Crowd motion pattern segmentation can macroscopically describe the holistic moving structures of crowds and simplify complex interactions among individuals to closely watch crowds with similar motion states. It does not only depict the segmentation in the spatial space but also reflects the motion tendency over a certain period. These patterns can be joint or disjoint in the image space [25]. The technique is regarded as an indispensable foundation for other crowd behaviour analysis techniques [22,26,27,28] and, therefore, received a lot of attention. However, improving the accuracy of the segmentation results is quite challenging, particularly resulting from confused crowd scenarios, high people density, low resolution, etc.Based on the principle of crowd motion detection [4,29], the existing methods of crowd motion segmentation and clustering can be classified into three main categories, including the flow field model-based [30,31,32], probability model-based [33,34,35], and similarity-based methods [36,37,38,39,40]. The first category uses flow field models to simulate image spatial segmentation and produce spatially continuous segments consequently. This type of method has been successful in dealing with high-density scenes, but it may result in creating over-segmented scenes in low-crowd-density scenes. The other two categories utilize local motion features by initially extracting them, then segmenting crowds using a variety of well-developed clustering algorithms. Their detection results are usually erratic, but these methods can be applied to structured and unstructured various crowd scenes. More specifically, the similarity-based clustering methods extract trajectories or tracklets as motion features and utilize similarity measurements for motion clustering or crowd profiling. This method has become more and more popular due to its virtually unsupervised process. It also has the advantage of being suitable for structured and unstructured scenes with different crowd-level density degrees. However, as the density of the crowd increases, the scene clutter becomes more severe. Feature extraction and tracking cannot be carried out accurately due to the problems of severe occlusion and clutter [41].In the available literature, many similarity-based clustering methods adopted keypoint trajectories obtained by the Kanade–Lucas–Tomasi (KLT) tracker [42,43] as their basis to describe the raw motion of the crowd data due to its robustness and computational efficiency [36,44]. This feature point tracker is often used as part of a larger tracking framework and the resulting sub-trajectories are a description of the microscopic behaviour of the crowd motion. It has many data points over a short time. They are compact spatiotemporal representations of moving rigid points [45]. In the subsequent step, these methods apply Coherent Neighbour Invariance (CNI) on KLT points for crowd motion clustering. It is worth mentioning that they use the KLT keypoints as raw input without making enhancements as pre-processing for their clustering techniques.In particular, the fragment of a trajectory obtained by the KLT tracker within a short range is called a tracklet. The length of the KLT tracklets of motion crowd depends on several factors. The most important factors are: (a) the frame rate of the frame sequence, (b) the relative position of the camera, and (c) the intensity of the motion patterns present in the scene. Furthermore, the kinematic sequence of a single tracklet’s points is assumed to be in a homogeneous localization in a consecutive time. All these factors can be exposed based on the location of moving points related to trajectories in a two-dimensional space. In some motion cases, however, the moving dense points of a single tracklet across frames are frequently lost in crowded scenes. More importantly, this indicates that many tracklet feature points can be lost across a few frames. Nonetheless, some of them are detected and tracked again in a few frames. Consequently, this condition has negatively affected the outcome of the crowd motion clustering and produced inaccurate results due to the lack of moving point information during the clustering process. Naturally, tracklets belonging to the same feature point should be merged into a single trajectory for more accurate motion crowd detection.Many previous studies on similarity-based coherent motion and crowd detection used the KLT tracker to create short trajectories as initial input data. The relationships of the moving keypoints as the main phase in the clustering process were analysed. However, previous studies focused on improving the tracklet keypoint clustering technique rather than the tracklet feature itself. The input feature development as a key factor in improving motion crowd clustering has been neglected in previous research. Therefore, this study aims to characterize disconnected KLT features, which mostly occur in the extracted input trajectories. To solve these problems, a tracklet-before-clustering initialization strategy is proposed to enhance coherent motion filtering. To the researchers’ best knowledge, this study is original in the field of computer vision as it aims to investigate comprehensively and systemically the instability of extracted moving input features (tracklets) from the vision’s perspective to achieve better coherent motion detection. The main contributions of this work can be summarized as follows:A Hierarchical Tracklet Association (HTA) algorithm is proposed as an initialization strategy to optimize coherent motion clustering. The purpose of the proposed framework is to address the disconnected tracklets problem of the input KLT features and carry out proper trajectories repair to enhance the performance of motion crowd clustering. In other words, HTA can be described as an enhanced initialization strategy for tracklet-before clustering.The coherent motion clustering results of the crowd were comprehensively examined and analysed on a crowd dataset, which is openly available to the public and contains a huge number of video clips.The rest of this paper is outlined as follows: several related works are presented in Section 2. Section 3 introduces the fundamentals of coherent motion filtering detection based on coherent neighbour invariance. Section 4 provides details of the proposed hierarchical tracklet association algorithm. Section 5 presents the evaluation metrics. Section 6 provides the conducted experiments, the findings of the study, and comparisons of several videos. Section 7 provides the conclusion. 2. Related WorksDue to surveillance application demands, crowd analysis has captivated researchers over the past decade. Detecting collective movements in crowd scenes is one of the hot topics in video surveillance. Researchers [46,47,48] found that crowds tend to form when many people exhibit similar motion patterns. Crowd applications vary depending on the use of handcrafted features to deep learning methods. To understand more about deep learning methods on crowd analysis, these recent surveys should be considered [4,20,49,50]. This section provides an overview of similarity-based clustering methods, which utilise the trajectories’ pattern recognition for crowd detection.Among numerous efforts, which were carried out to investigate this topic, many methods examined the Coherent Neighbour Invariance (CNI) concept on the motion of KLT keypoints and developed it from their point of view. The CNI concept is first introduced by Zhou et al. [36] to detect crowds with coherent motions from clutters by applying the Coherent Filtering (CF) method. CF utilizes spatial–temporal information and motion correlations to segment crowds over a short period. This input information is a set of moving tracklets detected by the KLT feature tracker and used to form motion groups. CNI has become a universal prior knowledge in collective scenes and is widely used to solve the problems of time series data clustering, such as crowd behaviour analysis [37,38,51]. In the same vein, Shao et al. [37] introduced group profiling to understand the group-level dynamics and properties. They first discovered the Collective Transition Prior (CT) from the initial CF clustering results obtained from [36]. The group collective transition prior is learned through EM iteration. Then, visual descriptors were provided to quantify intra- and inter-group properties, which were used for crowd detection and analysis. Chen et al. [52] proposed a Patch-based Topic Model (PTM) for group detection. The process begins with dividing the input crowd image into a fixed number of patches (using a Simple Linear Iterative Clustering algorithm). Then, a patch-level descriptor is computed for each patch by combining the feature points generated by the KLT tracker and the orientation distribution of each feature point within the patch. The Latent Dirichlet Allocation-based model is then combined with the Markov Random Field to determine the groups. Pai et al. [53] proposed a Spatio-Angular Density-based Clustering approach to cluster the crowd motion based on angular and spatial data obtained from the input trajectories. The data depends on the KNN similarity measure and angular deviation between the moving keypoints. Their work is effective to scene change when the tracks are well extracted. Wang et al. [54] proposed a self-weighted multi-view clustering approach that combines an orientation-based graph and a structural context-based graph. They applied a tightness-based merging strategy to detect groups within the crowd.Another similarity-clustering approach is proposed by Zhou et al. [44] for detecting KLT motions using the Crowd Collectiveness (MCC) descriptor. Collectiveness describes the degree to which the individuals act as a union in a collective motion. It depends on multiple factors, such as the decision-making of individuals and crowd density. First, the algorithm measured the collectiveness of each point by using graph-based learning. Then, it detected collective motions by thresholding the crowd collectiveness. It can be used to detect collective motions at different length scales from randomly moving outliers. However, the algorithm is sensitive when there is a break in paths and returns a large number of tracklets. Meanwhile, Shao et al. [55] proposed the collectiveness descriptor based on CT to detect and quantify the collectiveness of all group members. However, relying on motion attributes without refining showed irrelevant measurements of collectiveness. Japar et al. [56] proposed a discriminative visual-attributes extraction approach based on still-image input to detect the collective motion of the crowd. They classified individuals by head pose to infer individual-level collectiveness analysis, including collectiveness detection. Most of the above methods detect the collective motion in the crowd by considering moving keypoint relations as the main stage for the clustering process. However, researchers neglected to deal with the input feature development as an important factor in improving motion crowd clustering.This work differs significantly compared with the previous studies. This study focuses on the instability of extracted moving input features (tracklets) from the vision’s perspective. This study aims to utilize the short path of the input trajectories to enhance it as a correction strategy before clustering for further coherent motion enhancement. 3. The Fundamental of Coherent Filtering (CF) ClusteringWhen feature keypoints are used to describe objects in scenes, the process of the crowd motion analysis can be converted to analyse the motion states of these keypoints. CF is proposed by Zhou et al. [36]. It is a similarity-based clustering technique, which detects the crowd’s coherent motion from crowd clutters. A general illustration of the CF process is shown in Figure 1. It functions from a microscopic-to-macroscopic consistency based on two main sequence processes that are briefly described in Section 3.1 and Section 3.2. 3.1. Coherent Motion Cluster Detection Based on Coherent Neighbor Invariance (CNI)The coherent neighbour invariance, which is widely existing in collective scenes, can understand coherent motion detection. This section provides the details of the CNI relationship and describes its basic equations. It also provides the local neighbourhood of the moving individuals or points in terms of: (1) the invariance of spatiotemporal relationships of the points; and (2) the invariance of velocity correlations of points. The CNI process is shown in Figure 1a. It is categorized into two main sequence stages that are briefly described in Section 3.1.1 and Section 3.1.2. 3.1.1. Spatio-Temporal Invariant Points by Using Euclidian Distance Metric StageInitially, let the represented input moving points that contain all mixed points as I , which are the coherent and uncoherent moving points. All the points are moving in 2-dimensional space during a period from t to t + d , where d is the time distance. Here, the points are a general term, which represents the moving points. To find the spatiotemporal invariant points, each point from the input points must be tested separately from all other points. Therefore, at time t , select randomly point p , where p ∈ I and calculate the distance between point p and other point(s) p k , where p k ∈ I and p k ≠ p . For the calculation of the distance among the entire points, the Euclidean distance measure function is used for this purpose. It determines the quantitative degree of how close two points are. Then, find the K nearest points to point p and create the K nearest neighbours (KNN) group Ν t p at time t . Accordingly, the process is performed for point p to the rest of the specified period until t + d to search for the KNN point set Ν τ p , where τ = t → t + d . Finally, a large graph is created to filter out points that do not satisfy the neighbourhood condition, then search for the invariant neighbour among the KNN of point p from t to t + d , and represent it as M τ p . 3.1.2. Velocity Correlated Invariant Points StageTo identify the spatiotemporal invariant points, each point from the invariant neighbour set must be tested separately from all other points. Suppose the point p k belongs to the invariant neighbour set of point p , where each p k ∈ M τ p . Thus, at time t , compute the velocity correlation between p k and p . Second, compute the average velocity correlation g τ p k from time t to t + d with point p . Then, detect the outlier points from the invariant neighbour set M τ p based on the g τ p k

【本文地址】

公司简介

联系我们