Enhancing Security : Deep Learning Models for Anomaly Detection in Surveillance Videos
Karunarathne, Lihini (2024)
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:amk-2024053119175
https://urn.fi/URN:NBN:fi:amk-2024053119175
Tiivistelmä
An extensive study on effective methods to detect anomalous events in surveillance videos is presented in this research. An increasing demand for reliable and efficient anomaly detection systems capable of identifying various anomalies in real time was realized. This anomaly detection would further improve the lives of people by increasing security and safety.
To provide examples of how to enable real-time anomaly detection, this work explored and evaluated three different deep learning architectures for multi-class classification with residual networks and developed the models: 2DCNN+RNN, 3DCNN, and (2+1)DCNN. Then, the best-performing model was optimized with event excerpting techniques to enhance its performance in detecting and classifying anomalies with temporal context.
By analyzing the results, the followed method was effective in accurately identifying and classifying anomalous events in surveillance video. The optimized (2+1)DCNN Model exhibited superior performance. and achieved the Area Under the Curve (AUC) value of 81% on the dataset of University of Central Florida (UCF) Crime.
This study encourages using the fine-tuned optimized deep learning models for enhanced security and safety measures in a variety of surveillance environments and emphasizes the importance of high-quality video data and effective event excerpting techniques in maximizing the performance of deep learning models.
To provide examples of how to enable real-time anomaly detection, this work explored and evaluated three different deep learning architectures for multi-class classification with residual networks and developed the models: 2DCNN+RNN, 3DCNN, and (2+1)DCNN. Then, the best-performing model was optimized with event excerpting techniques to enhance its performance in detecting and classifying anomalies with temporal context.
By analyzing the results, the followed method was effective in accurately identifying and classifying anomalous events in surveillance video. The optimized (2+1)DCNN Model exhibited superior performance. and achieved the Area Under the Curve (AUC) value of 81% on the dataset of University of Central Florida (UCF) Crime.
This study encourages using the fine-tuned optimized deep learning models for enhanced security and safety measures in a variety of surveillance environments and emphasizes the importance of high-quality video data and effective event excerpting techniques in maximizing the performance of deep learning models.