Scholarlink

<p>The surge in illegal activities has underscored the critical role of surveillance cameras in monitoring and recording suspicious actions. However, the vast data generated by these systems poses a significant challenge in classifying abnormal activities. To address this, DReLNet is proposed for Abnormal Human Activity Classification in Theft Surveillance Video Using Reinforcement Learning-Based Fusion. The approach begins by extracting frames from surveillance videos, followed by noise reduction using a Stacked Bilateral Median Filter. Keyframes are identified via Absolute Consecutive Histogram Difference, and features are extracted using Capsule Net, DenseNet 201, and ResNet-152 models. These features are then fused, and abnormal activities are classified using the deep reinforcement learning-assisted model, DReLNet. The proposed model achieves an accuracy of 99.55% on the Theft dataset, significantly outperforming existing models such as DeepCNN, conv_LSTM, Bi_GRU and GRU. Higher accuracy enhances the model's overall performance by reducing the probability of incorrect classifications, which is critical for effective real-time surveillance systems.</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

DReLNet: abnormal human activity classification in theft surveillance video using reinforcement learning based fusion

  • 期刊论文

  • 发表日期：

    2025年1月17日

查看全文
  • Shubhangi Prabhakar Kale,
  • Shriram Raghunathan

摘要

The surge in illegal activities has underscored the critical role of surveillance cameras in monitoring and recording suspicious actions. However, the vast data generated by these systems poses a significant challenge in classifying abnormal activities. To address this, DReLNet is proposed for Abnormal Human Activity Classification in Theft Surveillance Video Using Reinforcement Learning-Based Fusion. The approach begins by extracting frames from surveillance videos, followed by noise reduction using a Stacked Bilateral Median Filter. Keyframes are identified via Absolute Consecutive Histogram Difference, and features are extracted using Capsule Net, DenseNet 201, and ResNet-152 models. These features are then fused, and abnormal activities are classified using the deep reinforcement learning-assisted model, DReLNet. The proposed model achieves an accuracy of 99.55% on the Theft dataset, significantly outperforming existing models such as DeepCNN, conv_LSTM, Bi_GRU and GRU. Higher accuracy enhances the model's overall performance by reducing the probability of incorrect classifications, which is critical for effective real-time surveillance systems.