Publication

Abstract : Video anomaly detection for crimes involves using machine learning algorithms to identify unusual or abnormal activities in surveillance footage, helping to automatically flag potential criminal behavior for further investigation. Liquid Neural Networks (LNNs) are “on-the-fly” brain-inspired models for time series. They adapt in real-time, unlike static networks, and have a dynamic structure that grows or shrinks with new data. This continuous learning and flexibility make them ideal for unpredictable data streams, leading to better predictions and smaller, faster models for tasks like robot control or anomaly detection. Multiple Instance Learning (MIL) with neural networks involves training a model using bags of instances, where each bag contains multiple instances, and only the collective label for the bag is known. Neural networks are employed to learn and classify based on the features of instances within each bag, making MIL suitable for tasks with weakly labeled or ambiguous data. The UCF Crime Dataset is a video dataset containing various crime scenarios, such as theft and assault, captured in an uncontrolled environment. It serves as a resource for developing and evaluating video-based crime detection and anomaly recognition algorithms. We use ROC AUC as our main metric along with common classification metrics like accuracy and precision. The Closed-form continuous-time neural network model with multiple instance learning (MIL), trained on I3D features, effectively classifies anomaly videos with high accuracy, precision, and ROC AUC scores, distinguishing between anomaly and normal videos. MIL leverages weakly labeled data, enhancing generalization and robustness, though real-time anomaly detection may pose challenges. Further refinement of the CF-CTNN model and its parameters could improve performance, highlighting its promising potential for video anomaly identification.