1,721,021 research outputs found
Visual Indoor Localization in Known Environments
Full text linkIn this letter, we propose a visual indoor localization technique, which localizes a camera sensor by comparing the acquired images to a reference model of location-tagged visual features. The proposed method relies on an efficient way to search for feature matches, which can run in real time. Experimental results show good localization accuracy even in challenging scenarios
On-line trajectory clustering for anomalous events detection
No full textIn this paper, we propose a trajectory clustering algorithm suited for video surveillance systems. Trajectories are clustered on-line, as the data are collected, and clusters are organized in a tree-like structure that, augmented with probability information, can be used to perform behaviour analysis, since it allows the identification of anomalous events. (c) 2006 Elsevier B.V. All rights reserved
Surveillance-Oriented Event Detection in Video Streams
No full textThe field of computer vision covers a large number of research topics, ranging from low-level processing aspects up to high-level image and video interpretation problems. This article gives a short introduction to security-oriented event analysis systems, whose aim is to give a semantic interpretation to video sequences in order to detect anomalous, dangerous or forbidden situations. The two main approaches to event analysis are here described, highlighting their advantages and limits. For each approach, a practical example is given, showing how events can be explicitly recognized in terms of their structure, or alternatively how they can be classified according to their degree of anomaly
Detecting moving people in video streams
No full textThe detection of moving people is an important task for video surveillance systems. This paper presents a motion segmentation algorithm for detecting people moving in indoor environments. The proposed algorithm works with mobile cameras and it is composed of two main parts. In the first part, a frame-by-frame procedure is applied to compute the difference image, and a neural network is used to classify whether the resulting image represents a static scene or a scene containing mobile objects. The second part tries to reduce the detection errors in terms of both false or missed alarms. A finite state automaton has been designed to give a robust classification and to reduce the number of false or missed blobs. Finally, a bounding ellipse is computed for each detected blob in order to isolate moving people. (c) 2005 Elsevier B.V. All rights reserved
Distributed Signature Fusion for Person Re-Identification
No full textIn many surveillance tasks it is very important for security operators to know whether a specific person is present in a given scene, at a given position and time. Person rei-dentification deals with this problem in order to provide more efficient security. A novel distributed appearance-based method for person re-identification is proposed. Spatio-temporal features
are used to group the camera network into camera neighbourhoods.
A intra-neighbourhood camera confidence hand-over measure
is computed by exploiting a signatures’ distance measure.
The camera confidence measure is exploited to save network
resources. Features that capture the chromatic appearance and the shape of an individual are used to compute a discriminative signature. The Expectation Maximization algorithm is used to fit Gaussian Mixture Models over the chromatic features. GMMs are exploited to compute the distance between signatures and to update the intra-neighbourhood camera confidence. The method
has been validated using a benchmark dataset and a new dataset acquired from a wide camera network scenario
- …
