Scholarlink

<p>Theft is a growing global concern, and while video surveillance systems are widely used for prevention, they rely on manual monitoring, which is inconsistent and labor-intensive. This study proposes a transfer learning approach for automatic theft detection using pre-trained human action recognition (HAR) models. Unlike prior works relying on handcrafted features or generic anomaly detection, we leverage spatio-temporal representations from pre-trained action recognition models to identify subtle theft behaviors, even with limited theft-specific data. Datasets were derived from the UCF-Crime dataset. Three balanced binary classification datasets were constructed, aggregating theft-related activities under a single "Theft" label against "Normal" behavior: the NS dataset (200 videos: 100 Normal vs. 100 Theft, from stealing), the NSS dataset (300 videos: 150 Normal vs. 150 Theft, from stealing and shoplifting), and the NSSR dataset (600 videos: 300 Normal vs. 300 Theft, from stealing, shoplifting, and robbery). Each video was split into 64-frame segments, resized to 224<InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\times \)</EquationSource> <EquationSource Format="MATHML"><math> <mo>×</mo> </math></EquationSource> </InlineEquation>224 pixels, and processed using pre-trained Inflated 3D ConvNet (I3D) and SlowFast models, extracting 400-dimensional class probability logits averaged into a single feature vector per video, yielding 12 derived feature sets. Binary classification was performed using Support Vector Machines, Decision Trees, Neural Networks, Random Forest, K-Nearest Neighbors, Gaussian Naive Bayes, and Gradient Boosting with hyperparameter tuning. The Neural Network model achieved the best performance with an F1-score of 0.90, Area Under the Curve of 0.90, accuracy of 0.90, recall of 0.91, and precision of 0.90. This study demonstrates how transfer learning can improve theft detection in surveillance systems, reducing reliance on human intervention. Future work will focus on real-time deployment feasibility.</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

A transfer learning approach to automatic theft detection

  • 期刊论文

  • 发表日期：

    2026年4月13日

查看全文
  • İrem Karaca Uluoğlu,
  • Barış Ethem Süzek

摘要

Theft is a growing global concern, and while video surveillance systems are widely used for prevention, they rely on manual monitoring, which is inconsistent and labor-intensive. This study proposes a transfer learning approach for automatic theft detection using pre-trained human action recognition (HAR) models. Unlike prior works relying on handcrafted features or generic anomaly detection, we leverage spatio-temporal representations from pre-trained action recognition models to identify subtle theft behaviors, even with limited theft-specific data. Datasets were derived from the UCF-Crime dataset. Three balanced binary classification datasets were constructed, aggregating theft-related activities under a single "Theft" label against "Normal" behavior: the NS dataset (200 videos: 100 Normal vs. 100 Theft, from stealing), the NSS dataset (300 videos: 150 Normal vs. 150 Theft, from stealing and shoplifting), and the NSSR dataset (600 videos: 300 Normal vs. 300 Theft, from stealing, shoplifting, and robbery). Each video was split into 64-frame segments, resized to 224 \(\times \) × 224 pixels, and processed using pre-trained Inflated 3D ConvNet (I3D) and SlowFast models, extracting 400-dimensional class probability logits averaged into a single feature vector per video, yielding 12 derived feature sets. Binary classification was performed using Support Vector Machines, Decision Trees, Neural Networks, Random Forest, K-Nearest Neighbors, Gaussian Naive Bayes, and Gradient Boosting with hyperparameter tuning. The Neural Network model achieved the best performance with an F1-score of 0.90, Area Under the Curve of 0.90, accuracy of 0.90, recall of 0.91, and precision of 0.90. This study demonstrates how transfer learning can improve theft detection in surveillance systems, reducing reliance on human intervention. Future work will focus on real-time deployment feasibility.