| dc.description.abstract | Behavior recognition and suspicious event analysis is a fundamental and important problem which can be applied to various applications like video surveillance, navigation, content-based image retrieval and so on. Its goal is to find the abnormal event of the environment and abnormal behavior no matter how the environmental conditions change.
This thesis proposes a novel method to detect carried objects from videos and applies it for suspicious event analysis, and presents a novel segmentation algorithm to segment a body posture into different body parts by using the technique of deformable triangulation. We discuss these two methods separately. In suspicious event analysis, First of all, a background subtraction using a minimum filter is proposed for detecting foreground objects. Then, a novel kernel-based tracking method is described for tracking each moving object and further obtaining its trajectory. With the trajectory, a novel ratio histogram is then proposed for analyzing the interactions between the carried object and its owner. After color re-projection, different carried objects can be accurately segmented from the background by taking advantages of GMMs (Gaussian mixture models). After bag detection, we propose an event analyzer to analyze various suspicious events using finite state machines. Even though there is no prior knowledge (like shape or color) about the bag, our proposed method still performs well to detect suspicious events from videos. As we know, due to the uncertainties of bag shape and color, there is no automatic system which can analyze various suspicious events (like robbery) caused by bags without any manual effects. However, by taking advantages of our proposed ratio histogram, different carried bags can be well segmented from videos and applied for event analysis. In Behavior recognition, First of all, to better analyze each posture, we triangulate it into triangular meshes, from which a spanning tree can be found using a depth-first search scheme. Then, two hybrid methods, i.e., the skeleton-based and model-driven ones, are proposed for segmenting the posture into different body parts according to its self-occlusion conditions. To analyze the self-occlusion condition, a novel clustering scheme is then proposed for clustering the training samples into a set of key postures. Then, a model space can be formed and used for posture classification and segmentation. After clustering, if the input posture belongs to the non-self-occlusion category, the skeleton-based scheme will be used for dividing it into different body parts which will be then refined using a set of Gaussian mixture models (GMMs). As to the self-occlusion case, we propose a model-driven technique for selecting a good reference model for guiding the process of body part segmentation. However, if two postures’ contours are similar, some ambiguity will be caused and lead to the failure in model selection. Thus, this thesis proposes a tree structure via a tracking technique for tackling this problem so that the best model can be selected not only from the current frame but also its previous frame. Thus, a suitable GMM-based segmentation scheme can be driven for finely segmenting a body posture into different body parts. Experimental results have proved that the proposed method is robust, accurate, and powerful in carried object detection and suspicious event analysis and in body part segmentation.
| en_US |