Automated Video Summarization for Suspicious Event Detection in Surveillance Systems: A Pipeline Approach
Syed Muhammad Hussain, Muhammad Azeem Haider, Mohammad Affan Ullah Habib, Muhammad Farhan · 2023
The increasing use of surveillance cameras to mitigate anomalous events has necessitated the development of automated methods for efficient video analysis and prompt detection of suspicious events. This paper presents a novel approach for automated video summarization utilizing the multi-stage Pipeline technique to identify suspicious events effectively. The proposed method extracts visual features from the input video as discrete events and subsequently classifies these events. Traditional surveillance systems rely on manual monitoring, which is laborintensive and prone to human error. The paper's approach focuses on using the OpenCV library to summarise the video. The model created for suspicious activity detection is created using the ConvLSTM and LRCN. The frames are divided into 14 sub-classes divided into normal and then thirteen other anomalous events. This classification process enables identifying suspicious events from normal occurrences, contributing to an accurate and reliable detection system. The model was trained using the publicly available DCSASS dataset and then tested on a custom-created dataset consisting of local videos from Pakistan to test its effectiveness in the local context. The results demonstrate that our method provides higher accuracy for both summarization quality and accuracy in detecting anomalous events. By reducing manual efforts and enhancing event detection capabilities, our approach can contribute to the proactive prevention and swift response to security threats, ultimately creating safer environments for society.