Scholarlink

<p>The burgeoning field of crowd behavior analysis necessitates advanced methodologies to ensure public safety and enhance event management strategies. Traditional approaches often fall short in addressing the complexity and dynamics of large crowd gatherings, leading to inaccuracies in density estimation, individual tracking, and behavior analysis. In response to these limitations, this research introduces a comprehensive suite of algorithms designed to revolutionize crowd behavior analysis through improved accuracy and adaptability. Firstly, we propose a dynamic density estimation algorithm integrating Faster R-CNN with Kernel Density Estimation, tailored for real-time adaptation to fluctuating crowd densities. This method not only refines person counting but also offers detailed spatial insights into crowd distribution, demonstrating a Mean Absolute Error of less than 5% when compared to ground truth data. Addressing the challenges in crowd tracking, we present a novel multi-object tracking framework that synergizes YOLOv8 object detection with the Hungarian algorithm to mitigate occlusions and trajectory ambiguities. This model maintains high tracking accuracy, achieving an Identity F1 score of over 0.8 even in densely populated environments. In the realm of crowd behavior recognition, our research introduces a hierarchical framework that amalgamates 3D Convolutional Neural Networks and Graph Convolutional Networks to dissect and understand both individual actions and collective behaviors. This innovative approach not only deciphers complex crowd dynamics but also sets a new benchmark in behavior recognition accuracy, exceeding 90% on UCSD Crowd and Avenue Dataset Samples. Furthermore, we pioneer a self-supervised abnormality detection framework leveraging Variational Adversarial Autoencoders, which circumvents the limitations of traditional anomaly detection by eliminating the need for labeled data samples. This model showcases remarkable adaptability and scalability, achieving an Area Under the Receiver Operating Characteristic Curve exceeding 0.9. Lastly, our multimodal emotion analysis framework, integrating Convolutional Neural Networks and Long Short-Term Memory networks, represents a significant leap forward in assessing collective emotional states, achieving an F1 score exceeding 0.85 for predominant emotions within crowds in different scenarios. Collectively, these advancements not only surpass existing methodologies in precision and adaptability but also offer new dimensions in crowd management and safety protocols.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Design of an iterative method for crowd behavior analysis integrating faster R-CNN, YOLOv8, and graph convolutional networks

  • 期刊论文

  • 发表日期：

    2025年5月12日

查看全文
  • Lokesh Heda,
  • Parul Sahare

摘要

The burgeoning field of crowd behavior analysis necessitates advanced methodologies to ensure public safety and enhance event management strategies. Traditional approaches often fall short in addressing the complexity and dynamics of large crowd gatherings, leading to inaccuracies in density estimation, individual tracking, and behavior analysis. In response to these limitations, this research introduces a comprehensive suite of algorithms designed to revolutionize crowd behavior analysis through improved accuracy and adaptability. Firstly, we propose a dynamic density estimation algorithm integrating Faster R-CNN with Kernel Density Estimation, tailored for real-time adaptation to fluctuating crowd densities. This method not only refines person counting but also offers detailed spatial insights into crowd distribution, demonstrating a Mean Absolute Error of less than 5% when compared to ground truth data. Addressing the challenges in crowd tracking, we present a novel multi-object tracking framework that synergizes YOLOv8 object detection with the Hungarian algorithm to mitigate occlusions and trajectory ambiguities. This model maintains high tracking accuracy, achieving an Identity F1 score of over 0.8 even in densely populated environments. In the realm of crowd behavior recognition, our research introduces a hierarchical framework that amalgamates 3D Convolutional Neural Networks and Graph Convolutional Networks to dissect and understand both individual actions and collective behaviors. This innovative approach not only deciphers complex crowd dynamics but also sets a new benchmark in behavior recognition accuracy, exceeding 90% on UCSD Crowd and Avenue Dataset Samples. Furthermore, we pioneer a self-supervised abnormality detection framework leveraging Variational Adversarial Autoencoders, which circumvents the limitations of traditional anomaly detection by eliminating the need for labeled data samples. This model showcases remarkable adaptability and scalability, achieving an Area Under the Receiver Operating Characteristic Curve exceeding 0.9. Lastly, our multimodal emotion analysis framework, integrating Convolutional Neural Networks and Long Short-Term Memory networks, represents a significant leap forward in assessing collective emotional states, achieving an F1 score exceeding 0.85 for predominant emotions within crowds in different scenarios. Collectively, these advancements not only surpass existing methodologies in precision and adaptability but also offer new dimensions in crowd management and safety protocols.