This research overcomes a critical limitation in surveillance AI by using Large Vision-Language Models to recognize actions in security footage without requiring labeled training data.

• Addresses the challenge of limited surveillance datasets by using zero-shot capabilities of LVLMs
• Demonstrates effectiveness despite typical surveillance challenges (poor viewpoints, low quality footage)
• Enables detection of suspicious activities without extensive model fine-tuning
• Offers a practical solution for expanding AI surveillance with reduced data requirements

This advancement matters for security applications by allowing rapid deployment of intelligent monitoring systems in new environments without collecting massive training datasets first.

Zero-Shot Action Recognition in Surveillance Videos