Real-Time Unattended Package Detection Using a Syntactic Approach

Abstract

Detecting unattended packages in video is a challenging surveillance problem. The goal is to detect packages and their owners, determine relationships between them and the environment, and recognize when an owner abandons a package. Concurrent events and complex interactions create problems for existing motion-based systems. Errors in target detection and tracking caused by shadows, noise, and occlusions create additional problems. We present a real-time system that addresses these issues and recognizes the unattended package event syntactically using a stochastic attribute grammar. Our system can detect events that occur concurrently, are corrupted by target and detection errors, or contain packages hidden behind other objects such as trash cans.

In this thesis, we review existing motion-based methods, then present our system and show results on unattended package detection in situations that would cause motion-based systems to fail.