Pawel Gardzinski

Human activity recognition in multiview video

In this paper, a novel multiview video based human activity recognition system which automatic detects of such behavior as fainting, a fight or a call for help is presented. The approach proposed in this paper used a directed graphical model based on propagation nets, a subset of dynamic Bayesian networks approaches, to model the behaviors. The performance of activity recognition is analyzed for three methods of characteristic points forming a behavior descriptor (four extreme points over contour, four extreme points over contour with different normalization process and n-evenly distributed points on the contour). The results prove high score of recognition of the system for “Calling for help”, “Faint”, “Fight”, “Falling” and “Bend at the waist” behaviors.

Human Behavior Recognition Using Negative Curvature Minima and Positive Curvature Maxima Points

Recently, automated human behavior recognition are studied in the context of many new applications such as content-based video annotation and retrieval, highlight extraction, video summarization and video surveillance. In this chapter a novel description of human pose - a combination of negative curvature minima (NCM) and positive curvature maxima (PCM) points are proposed. Experimental results are provided in the chapter in order to demonstrate precision of the human activity recognition versus size of the descriptor (a temporal interval durations between the nodes of the model). The experimental results are focused on recognition of call for help behavior. The results prove high score of recognition of the proposed method.

Unsupervised abnormal crowd activity detection in surveillance systems

We propose an unsupervised method for abnormal crowd activity detection in surveillance systems. Proposed solution is using MPEG-7 Motion Activity descriptors and Particle Filter algorithm for classification. The experiments were performed on UMN dataset sequences. The detection results are comparable to results obtained by supervised methods.

Human silhouette segmentation using discrete poisson equation and extended watershed algorithm

In this paper a novel approach on human silhouette segmentation for surveillance systems was proposed. The described solution uses discrete Poisson equation and a combination of extended watershed algorithm with Region Growing algorithm. Experiments were performed on a commonly known database PETS 2006 and the results show that the proposed solution achieves high precision and accuracy.

Nienadzorowana detekcja anormalnych zachowań tłumu w dozorowych sekwencjach wizyjnych

We propose an unsupervised method for abnormal crowd activity detection in surveillance systems. Proposed solution is using MPEG-7 Motion Activity descriptors and Particle Filter algorithm for classification. The experiments were performed on UMN dataset sequences. The detection results are comparable to results obtained by supervised methods. Słowa kluczowe: filtr cząsteczek, nienadzorowana detekcja anomalii, UMN

3-D Reconstruction of Real Objects Using an Android Device

In the paper, the autonomous system of reconstruction of 3-D model based on the matching characteristic features between the images for mobile devices with Android OS is proposed. Our method focuses on fully automated system with marker less calibration method. The experimental results show that although the reconstructed objects contain certain artifacts or loss, the end result can be successfully used by the average user.

Crowd density estimation based on voxel model in multi-view surveillance systems

In this paper, a novel crowd density estimation method based on voxel modeling in multi-view surveillance systems is presented. The approach proposed in this paper is based on human silhouette modeling with an anthropometric cylinder. The performance of crowd density estimation was analyzed on two multi-view sequences datasets. For this propose PETS 2006 and PETS 2009 were used. Performance of the proposed approach has been evaluated for two metrics: people counting and crowd classification.

Human Activity Interpretation Using Evenly Distributed Points on the Human Hull

In this paper, a human activity recognition system which automatically detects human behaviors in video is presented. The solution presented in this paper uses a directed graphical model with proposed by the authors Evenly Distributed Points (EDP) method. The experimental results prove efficient representation of the human activity and high score of recognition.