Scholarlink

<p>As the prevalence of video surveillance continues to rise dramatically, the development of an automated approach for detecting abnormal events within video footage becomes imperative. Numerous methods for anomaly detection in video surveillance have been suggested, with a recent trend focusing on leveraging deep learning techniques to achieve heightened accuracy in anomaly detection. However, a notable gap exists in the literature, as much of the recent research primarily emphasizes anomaly detection accuracy without giving due consideration to the equally crucial aspect of anomaly detection speed, particularly vital in the context of video surveillance. This research introduces a novel framework for anomaly detection called Depth-wise Deep Convolutional Anomaly Detection (DWDCAD). The necessity for a quick, precise, and effective anomaly detection framework in video surveillance that can operate using a lightweight convolutional architecture, multi-feature extraction, and optimum segmentation is what inspired the proposed model. The proposed model comprises four main phases: preprocessing, segmentation, feature extraction, and anomaly detection. Initially, the input video undergoes frame conversion, followed by the application of a Gaussian filtering approach is conducted to achieve image smoothing, noise reduction, and feature enhancement. The preprocessed image is then subjected to segmentation using a Modified Mask Region-CNN (MMR-CNN) scheme, which divides the image into meaningful segments. MMR-CNN incorporates three backbone frameworks with four Mask heads to obtain Regions of Interest (ROI). Subsequently, features are extracted from the segmented image, including Deep features (GoogleNet, InceptionV3), color features, Improved LGIP, and Hierarchical of skeleton (HOS). The Improved LGIP computes neighbor pixel values to extract relevant features. The extracted feature set is then fed into the Anomaly Detection model, utilizing a Depth-wise Deep Convolutional Neural Network (DW-DCNN) consisting of key blocks such as C-Conv and DS-Conv. Finally, the DW-DCNN model efficiently detects anomalies in the objects. The DW-DCNN scheme outperforms DCNN, DenseNet, MobileNet, ResNet, DCNN, ConvLSTM, and CASNET with the lowest MAE of 0.095 with a training data percentage of 60.</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

Modified convolutional architecture based anomaly detection for video surveillance with mask R-CNN segmentation model for object detection

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Gudupudi Ravikumar,
  • Malathy Vanniappan,
  • K. RajKumar

摘要

As the prevalence of video surveillance continues to rise dramatically, the development of an automated approach for detecting abnormal events within video footage becomes imperative. Numerous methods for anomaly detection in video surveillance have been suggested, with a recent trend focusing on leveraging deep learning techniques to achieve heightened accuracy in anomaly detection. However, a notable gap exists in the literature, as much of the recent research primarily emphasizes anomaly detection accuracy without giving due consideration to the equally crucial aspect of anomaly detection speed, particularly vital in the context of video surveillance. This research introduces a novel framework for anomaly detection called Depth-wise Deep Convolutional Anomaly Detection (DWDCAD). The necessity for a quick, precise, and effective anomaly detection framework in video surveillance that can operate using a lightweight convolutional architecture, multi-feature extraction, and optimum segmentation is what inspired the proposed model. The proposed model comprises four main phases: preprocessing, segmentation, feature extraction, and anomaly detection. Initially, the input video undergoes frame conversion, followed by the application of a Gaussian filtering approach is conducted to achieve image smoothing, noise reduction, and feature enhancement. The preprocessed image is then subjected to segmentation using a Modified Mask Region-CNN (MMR-CNN) scheme, which divides the image into meaningful segments. MMR-CNN incorporates three backbone frameworks with four Mask heads to obtain Regions of Interest (ROI). Subsequently, features are extracted from the segmented image, including Deep features (GoogleNet, InceptionV3), color features, Improved LGIP, and Hierarchical of skeleton (HOS). The Improved LGIP computes neighbor pixel values to extract relevant features. The extracted feature set is then fed into the Anomaly Detection model, utilizing a Depth-wise Deep Convolutional Neural Network (DW-DCNN) consisting of key blocks such as C-Conv and DS-Conv. Finally, the DW-DCNN model efficiently detects anomalies in the objects. The DW-DCNN scheme outperforms DCNN, DenseNet, MobileNet, ResNet, DCNN, ConvLSTM, and CASNET with the lowest MAE of 0.095 with a training data percentage of 60.