Anthony Stefanidis

Self-organised clustering for road extraction in classified imagery

Abstract The extraction of road networks from digital imagery is a fundamental image analysis operation. Common problems encountered in automated road extraction include high sensitivity to typical scene clutter in high-resolution imagery, and inefficiency to meaningfully exploit multispectral imagery (MSI). With a ground sample distance (GSD) of less than 2 m per pixel, roads can be broadly described as elongated regions. We propose an approach of elongated region-based analysis for 2D road extraction from high-resolution imagery, which is suitable for MSI, and is insensitive to conventional edge definition. A self-organising road map (SORM) algorithm is presented, inspired from a specialised variation of Kohonens self-organising map (SOM) neural network algorithm. A spectrally classified high-resolution image is assumed to be the input for our analysis. Our approach proceeds by performing spatial cluster analysis as a mid-level processing technique. This allows us to improve tolerance to road clutter in high-resolution images, and to minimise the effect on road extraction of common classification errors. This approach is designed in consideration of the emerging trend towards high-resolution multispectral sensors. Preliminary results demonstrate robust road extraction ability due to the non-local approach, when presented with noisy input.

Report from the first workshop on geo sensor networks

Advances in sensor technology and deployment strategies are revolutionizing the way that geospatial information is collected and analyzed. For example, cameras and GPS sensors on-board static or mobile platforms have the ability to provide continuous streams of geospatially-rich information. Furthermore, with the advent of nano-technology it becomes feasible and economically viable to develop and deploy low-cost, low-power devices that are generalpurpose computing platforms with multi-purpose on-board sensing and wireless communications capabilities. Special IT infrastructure challenges are posed by systems consisting of large numbers of unattended, untethered and collaborative sensor nodes that have small, non-renewable power supply and communicate via short range radio frequency with neighboring nodes. All these types of sensors may act collaboratively as nodes within broader network configurations. Such configurations may range in scale from few cameras monitoring traffic to thousands of nodes monitoring an ecosystem. The challenge of sensor networks is to aggregate sensor nodes into computational infrastructures that are able to produce globally meaningful information from raw local data obtained by individual sensor nodes. In geo sensor networks the geospatial content of the information collected, aggregated, analyzed, and monitored by a sensor network is fundamental; this might be performed locally in real-time on the sensor nodes or between sensor nodes, or off-line in a scattered or central repositories. Thus, a geosensor network may be loosely defined as a sensor network that monitors phenomena in a geographic space. This space may range in scale from the confined environment of a room to the highly complex dynamics of a an ecosystem region. The spatial aspect of the overall technology may be of importance in multiple levels of a geo sensor network, as the concepts of space, location, topology, and spatiotemporal events may be recognized on various abstraction levels. For example, the hardware and communication layers handle the physical space of sensor deployment, and communication topologies. The database layer generates execution plans for spatiotemporal queries that relate to sensor node location, and groups of sensors. Applications deal with the relation between sensor networks and phenomena in a geographic space. We feel that the academic and practical expertise of the spatial information theory and engineering domain are crucial to advance the development of sensor networks on all different abstraction levels. The ultimate objective is to develop generic sensor network programming infrastructure that is reusable, and widely applicable in all types of different domains.

An environment for content-based image retrieval from large spatial databases

In this paper, we address the problem of content-based image retrieval using queries on shape and topology. We focus on the particularities of image databases encountered in typical topographic applications and present the development of a spatial data management system that enables such queries. The query requires user-provided sketches of the shape and spatial configuration of the object (or objects) which should appear in the images to be retrieved. The objective of the search is to retrieve images that contain a configuration of objects sufficiently similar to the one specified in the query. Our approach combines the design of an integrated database with the development of a feature library and the necessary matching tools. In this paper, we present our overall scheme, introduce some individual database components, and provide some implementation results.

3D trajectory matching by pose normalization

Recent technological advances have made it possible to collect large amounts of 3D trajectory data. Such data play an essential role in numerous applications and are becoming increasingly important in mobile computing. One of the fundamental challenges in many of these application areas is the assessment of similarity between trajectories. As objects moving in a 3D space may often exhibit a similar motion pattern but may differ in location, orientation, and scale, the similarity assessment method employed must be invariant to these seven degrees of freedom. Previous work has addressed this problem primarily through local measures, such as curvature and torsion and has mostly concentrated on 2D trajectory data. This paper introduces a novel non iterative 3D trajectory matching framework that is translation, rotation, and scale invariant. We achieve this through the introduction of a pose normalization process that is based on physical principles, which incorporates both spatial and temporal aspects of trajectory data. We also introduce a new shape signature that utilizes the invariance that is achieved through pose normalization. The proposed scheme was tested both on simulated data and on real world data and has shown to offer improved robustness compared to local measures.

Summarizing video datasets in the spatiotemporal domain

We address the problem of analyzing and managing complex dynamic scenes captured in video. We present an approach to summarize video datasets by analyzing the trajectories of objects within them. Our work is based on the identification of nodes in these trajectories as critical points that describe the behavior of an object over a video segment. The time instances that correspond to these nodes are used to select critical frames for a video summary that describes adequately and concisely an objects behavior within a video segment. The analysis of relative positions of objects of interest within the video feed may dictate the selection of additional critical frames, to ensure the separability of converging trajectories. The paper presents a framework for video summarization using this approach, and addresses the use of self-organizing maps to identify trajectory nodes.

Automated Extraction of Linear Features from Aerial Imagery Using Kohonen Learning and GIS Data

An approach to semi-automated linear feature extraction from aerial imagery is introduced in which Kohonens self-organizing map (SOM) algorithm is integrated with existing GIS data. The SOM belongs to a distinct class of neural networks which is characterized by competitive and unsupervised learning. Using radiometrically classified image pixels as input, appropriate SOM network topologies are modeled to extract underlying spatial structures contained in the input patterns. Coarse-resolution GIS vector data is used for network weight and topology initialization when extracting specific feature components. The Kohonen learning rule updates the synaptic weight vectors of winning neural units that represent 2-D vector shape vertices. Experiments with high-resolution hyperspectral imagery demonstrate a robust ability to extract centerline information when presented with coarse input.

Modeling and comparing change using spatiotemporal helixes

Spatiotemporal helixes are a novel way to model spatiotemporal change. They represent both the movement of an object, as it is expressed by the trajectory of its center, and the changes of its outline. Accordingly they are highly suitable to communicate the evolution of phenomena as they are captured e.g. in sequences of imagery. In this paper we present the spatiotemporal helix model and introduce spatiotemporal similarity metrics for the comparison of helixes. These metrics allow us to compare the behavior of different objects over time, and express the degree of their similarity. To demonstrate the application of our models and metrics we present experimental results.

Efficient summarization of spatiotemporal events

Raster datasets at discrete temporal instances capture a variety of spatiotemporal phenomena. These phenomena and the respective datasets that capture them may span various spatial and temporal scales, like a cars trajectory as it is captured by various cameras at video rates, or the long-term erosion of a rivers edge captured by an annual series of aerial imagery.

Automated spatiotemporal scaling for video generalization

We present a technique for the summarization and spatiotemporal scaling of video content. A self organizing map (SOM) neural network can be used to acquire a rough generalization of the spatiotemporal trajectories of moving objects, in the form of few selected nodes along these trajectories. We introduce a hybrid technique, combining SOM with geometric analysis to properly densify these nodes, to better represent the spatiotemporal behavior of objects. This allows us to bypass problems inherently associated with parameter selection in SOM. We also demonstrate how spatiotemporal scaling supports the analysis of behavioral patterns. The paper shows that our novel technique is a powerful tool for the extraction of generalized information from complex trajectories, displaying high invariance to noise and information gaps in the video stream. Experimental results demonstrate the accuracy potential of our generalization technique.

Spatiospectral cluster analysis of elongated regions in aerial imagery

The extraction of road networks from digital imagery is a fundamental operation in geospatial applications. In images captured by new satellite sensors with a ground sample distance of less than 2 meters per pixel, roads can be broadly described as elongated regions. We introduce a novel technique of spatiospectral cluster analysis in which the spatial properties of elongated regions are identified from unsupervised analysis of their corresponding spectral properties. Preliminary results demonstrate a fully automated process in which road centerline topology can be identified in high-resolution aerial imagery in the presence of typical clutter.