Graph-based Approach for Anomaly Detection in Video Surveillance
With the rise of technology, cameras are found to be a very useful tool to monitor different areas for public safety and security. Human operators of modern surveillance systems are confronted with a huge amount of digital video data that needs to be monitored constantly. However, due to many limitations such as fatigue and time efficiency, human operators are not continuously watching these surveillance videos. Hence, they are not able to prevent incidents and reduce the scale of mishaps. In some cases, the recorded videos are watched after the incident to identify when it happened. However, given the massive amount of video data recorded and the number of cameras that might be involved, the identification of anomalies will take a significant amount of time. This manual analysis of video data requires extensive manpower and is lacking efficacy and reliability. This dissertation propose two anomaly detection systems to overcome these issues. An offline framework to detect anomalies after they happen and an online framework to identify incidents while they are happening to limit their damage. The contributions of this dissertation are divided into three main parts. The first part introduces a method based on dynamic graphs to represent motion and semantic information present in the scene. This method utilizes hand-crafted features as well as graphs properties to extract useful information from video sequences. The second part presents an offline framework that uses a one-class classifier to model normal activities represented by spatiotemporal features and features extracted using the graph-based approach. In the third part, we propose to learn regularities using autoencoders. This approach is based on the intuition that the autoencoder learns to reproduce the test input with good fidelity if it resembles the training data, which will produce low reconstruction errors for normal instances. Therefore, reconstruction error between the network’s output and the target output is used to classify a video sequence as normal or abnormal. Quantitative and qualitative experiments show that the proposed approaches can reach state- of-the-art performance on three challenging anomaly detection datasets that represent real-world scenarios.