Video Event Identification using Weakly Supervised Anomaly Detection

Sammanfattning

The increased use of surveillance cameras and the growing volume of collected data make it impractical for humans to manually review long videos to identify interesting events. A potential solution to this problem is the implementation of automatic video anomaly detection systems that automatically detect interesting segments. This thesis investigates the capability of a two-stage Multiple Instance Learning (MIL) model for anomaly detection and explores methods to enhance its performance and reduce inference time. Different feature extraction methods and pseudo-label refinement are strategies investigated to boost performance. The best-performing model, according to AUC-value, incorporates a Glance-Focus approach, utilizing two attention blocks in the first stage. This model achieved an AUC of 82.36% in the first stage and 83.69% in the second stage on the UCF-Crime dataset. This model also generated the second largest improvement in AUC between the stages. The results demonstrate that while refined feature extraction and pseudo-labeling can boost early-stage performance, this does not necessarily boost the second stage.