This research introduces a new benchmark for evaluating how effectively multimodal large language models can analyze surveillance footage for security applications.

• Addresses limitations in existing benchmarks by providing MLLM-style question-answering capabilities
• Focuses specifically on assessing models' abilities to understand anomalous events in video surveillance
• Enables more accurate evaluation of AI models' open-ended text responses to security scenarios

For security professionals, this benchmark represents a significant advancement in how we evaluate AI systems for real-world surveillance applications, potentially improving anomaly detection in critical security infrastructure.

A Benchmark for Crime Surveillance Video Analysis with Large Models