Human behavior analysis has become an active topic of great interest and relevance for a number of applications and areas of research. The research in recent years has been considerably driven by the growing level of criminal behavior in large urban areas and increase of terroristic actions. Also, accurate behavior studies have been applied to sports analysis systems and are emerging in healthcare. When compared to conventional action recognition used in security applications, human behavior analysis techniques designed for embedded applications should satisfy the following technical requirements: (1) Behavior analysis should provide scalable and robust results; (2) High-processing efficiency to achieve (near) real-time operation with low-cost hardware; (3) Extensibility for multiple-camera setup including 3-D modeling to facilitate human behavior understanding and description in various events. The key to our problem statement is that we intend to improve behavior analysis performance while preserving the efficiency of the designed techniques, to allow implementation in embedded environments. More specifically, we look into (1) fast multi-level algorithms incorporating specific domain knowledge, and (2) 3-D configuration techniques for overall enhanced performance. If possible, we explore the performance of the current behavior-analysis techniques for improving accuracy and scalability. To fulfill the above technical requirements and tackle the research problems, we propose a flexible behavior-analysis framework consisting of three processing-layers: (1) pixel-based processing (background modeling with pixel labeling), (2) object-based modeling (human detection, tracking and posture analysis), and (3) event-based analysis (semantic event understanding). In Chapter 3, we specifically contribute to the analysis of individual human behavior. A novel body representation is proposed for posture classification based on a silhouette feature. Only pure binary-shape information is used for posture classification without texture/color or any explicit body models. To this end, we have studied an efficient HV-PCA shape-based descriptor with temporal modeling, which achieves a posture-recognition accuracy rate of about 86% and outperforms other existing proposals. As our human motion scheme is efficient and achieves a fast performance (6-8 frames/second), it enables a fast surveillance system or further analysis of human behavior. In addition, a body-part detection approach is presented. The color and body ratio are combined to provide clues for human body detection and classification. The conventional assumption of up-right body posture is not required. Afterwards, we design and construct a specific framework for fast algorithms and apply them in two applications: tennis sports analysis and surveillance. Chapter 4 deals with tennis sports analysis and presents an automatic real-time system for multi-level analysis of tennis video sequences. First, we employ a 3-D camera model to bridge the pixel-level, object-level and scene-level of tennis sports analysis. Second, a weighted linear model combining the visual cues in the real-world domain is proposed to identify various events. The experimentally found event extraction rate of the system is about 90%. Also, audio signals are combined to enhance the scene analysis performance. The complete proposed application is efficient enough to obtain a real-time or near real-time performance (2-3 frames/second for 720×576 resolution, and 5-7 frames/second for 320×240 resolution, with a P-IV PC running at 3GHz). Chapter 5 addresses surveillance and presents a full real-time behavior-analysis framework, featuring layers at pixel, object, event and visualization level. More specifically, this framework captures the human motion, classifies its posture, infers the semantic event exploiting interaction modeling, and performs the 3-D scene reconstruction. We have introduced our system design based on a specific software architecture, by employing the well-known "4+1" view model. In addition, human behavior analysis algorithms are directly designed for real-time operation and embedded in an experimental runtime AV content-analysis architecture. This executable system is designed to be generic for multiple streaming applications with component-based architectures. To evaluate the performance, we have applied this networked system in a single-camera setup. The experimental platform operates with two Pentium Quadcore engines (2.33 GHz) and 4-GB memory. Performance evaluations have shown that this networked framework is efficient and achieves a fast performance (13-15 frames/second) for monocular video sequences. Moreover, a dual-camera setup is tested within the behavior-analysis framework. After automatic camera calibration is conducted, the 3-D reconstruction and communication among different cameras are achieved. The extra view in the multi-camera setup improves the human tracking and event detection in case of occlusion. This extension of multiple-view fusion improves the event-based semantic analysis by 8.3-16.7% in accuracy rate. The detailed studies of two experimental intelligent applications, i.e., tennis sports analysis and surveillance, have proven their value in several extensive tests in the framework of the European Candela and Cantata ITEA research programs, where our proposed system has demonstrated competitive performance with respect to accuracy and efficiency.
|Qualification||Doctor of Philosophy|
|Award date||12 Oct 2011|
|Place of Publication||Eindhoven|
|Publication status||Published - 2011|