果冻传媒

Inhabited areas all around the world witness various criminal activities ranging from misdemeanors, such as trespassing, to grave threats, such as intentional homicides and terrorist attacks. In order to respond to these threats, many public and private organizations invest heavily in trained workforce and technology to efficiently prevent, detect, and counteract criminal activities. Conventionally, safety was ensured by real-time camera surveillance and scene understanding performed by manual human interpretation. However, as the physical dimensions of the surveillance area and the number of cameras grow proportionally, the human interpretation of many concurrent video streams becomes infeasible for human supervision only, and parallel inspection is too expensive in terms of labor. This thesis of concentrates on computer vision and machine learning techniques that enable behavior analysis of scene participants for automated surveillance applications.

Surveillance first line-of-defense against crime

Surveillance is an act of close observation of an area-of-interest and it is the first line-of-defense against crime. Performing surveillance is useful before, during, and after a criminal activity. Prior to a criminal activity, surveillance is an efficient deterrent and a preventive measure. During a criminal activity, it provides information and means to intervene, while thereafter, it supplies vital information for law enforcement and justice systems. Therefore, surveillance is widely adopted all around the world to ensure the safety and security of people and to protect valuable commodities.

Technologies of behavior analysis

In this research, several key enabling technologies of behavior analysis are explored prior to outlining a scalable and holistic automated surveillance system design. The first part of this work explores the feature extraction stage of re-identification. The second part focuses on datasets for re-identification, including an investigation of the optimal data acquisition strategies and the open-set re-identification problem. Finally, the third part of the thesis concentrates on video anomaly detection. In addition, a scalable automated surveillance system design is proposed that combines re-identification and anomaly detection.

Proof-of-concept provides promising performance

A proof-of-concept prototype of the proposed framework has been installed in a public environment (Port of Moerdijk, The Netherlands) where security is a clear priority and addressing threats is important for smooth operation. The industrial scalability of the solution is studied as part of the Interreg PASSAnT Project. The proposed approach enables the detection of a wide-range of abnormal behaviors in real time. Furthermore, the proposed framework is evaluated for its performance in a real-world setting, and has shown to provide promising performance using both numerical and empirical experiments. Experimental results indicate that it is possible to automatically detect a broad range of anomalous behavior with an affordable initial investment cost.

The contributions of this thesis aim to pave the way for practical and high-performance realizations of automated surveillance thereby offering a feasible, scalable alternative to fully manual supervision.

Title of PhD thesis: . Promotor: Prof. Peter de With. Co-promotor: Associate Prof. Egor Bondarau.