Roland Miezianko

Incremental Multiple Kernel Learning for object recognition

A good training dataset, representative of the test images expected in a given application, is critical for ensuring good performance of a visual categorization system. Obtaining task specific datasets of visual categories is, however, far more tedious than obtaining a generic dataset of the same classes. We propose an Incremental Multiple Kernel Learning (IMKL) approach to object recognition that initializes on a generic training database and then tunes itself to the classification task at hand. Our system simultaneously updates the training dataset as well as the weights used to combine multiple information sources. We demonstrate our system on a vehicle classification problem in a video stream overlooking a traffic intersection. Our system updates itself with images of vehicles in poses more commonly observed in the scene, as well as with image patches of the background, leading to an increase in performance. A considerable change in the kernel combination weights is observed as the system gathers scene specific training data over time. The system is also seen to adapt itself to the illumination change in the scene as day transitions to night.

Detecting and recognizing abandoned objects in crowded environments

In this paper we present a framework for detecting and recognizing abandoned objects in crowded environments. The two main components of the framework include background change detection and object recognition. Moving blocks are detected using dynamic thresholding of spatiotemporal texture changes. The background change detection is based on analyzing wavelet transform coefficients of non-overlapping and non-moving 3D texture blocks. Detected changed background becomes the region of interest which is scanned to recognize various objects under surveillance such as abandoned luggage. The object recognition is based on model histogram ratios of image gradient magnitude patches. Supervised learning of the objects is performed by support vector machine. Experimental results are demonstrated using various benchmark video sequences (PETS, CAVIAR, i-Lids) and an object category dataset (CalTech256).

People detection in low resolution infrared videos

In this paper we present a method for detecting people in low resolution infrared videos. We further explore the feature set based on histogram of gradients beyond the well received HOG descriptors. Our approach is based on extracting gradient histograms from recursively generated patches and subsequently computing histogram ratios between the patches. Each set of patches is defined in terms of relative position within the search window, and each set is then recursively applied to extract smaller patches. The histogram of gradient ratios between patches become the feature vector. We adopted a linear SVM classifier as it provides a fast and effective framework for feature descriptor processing with minimal parameter tuning. Experimental results are presented on various OTCBVS datasets.

Localization of detected objects in multi-camera network

In this paper we present a framework for detecting, recognizing, and localizing objects in overlapping multi-camera network. The three main components of the framework include background change detection, object recognition, and object localization. The background change detection is based on analyzing wavelet transform coefficients of small patches of non-overlapping 3D texture maps. Detected changed background becomes the region of interest which is scanned to recognize various objects of interest. The object recognition is based on model histogram ratios of gradient magnitude patches. The supervised learning of objects is performed by a support vector machine. A homographic spatial transformation brings multiple cameras into alignment with the ground plane to localize objects in 2D space. Experimental results are demonstrated using various benchmark video sequences and object category datasets.

Texture Dissimilarity Measures for Background Change Detection

Presented framework provides a method for adaptive background change detection in video from monocular static cameras. A background change constitutes of objects left in the scene and objects moved or taken from the scene. This framework may be applied to luggage left behind in public places, to asses the damage and theft of public property, or to detect minute changes in the scene. The key elements of the framework include spatiotemporal motion detection, texture classification of non-moving regions, and spatial clustering of detected background changes. Motion detection based on local variation of spatiotemporal texture separates the foreground and background regions. Local background dissimilarity measurement is based on wavelet decomposition of localized texture maps. Dynamic threshold of the normalized dissimilarity measurement identifies changed local background blocks, and spatial clustering isolates the regions of interest. The results are demonstrated on the PETS 2006 video sequences.

Detection of suspicious activity using incremental outlier detection algorithms

Detection of unusual trajectories of moving objects can help in identifying suspicious activity on convoy routes and thus reduce casualties caused by improvised explosive devices. In this paper, using video imagery we compare efficiency of various techniques for incremental outlier detection on detecting unusual trajectories on simulated and real-life data obtained from SENSIAC database. Incremental outlier detection algorithms that we consider in this paper include incremental Support Vector Classifier (incSVC), incremental Local Outlier Factor (incLOF) algorithm and incremental Connectivity Outlier Factor (incCOF) algorithm. Our experiments performed on ground truth trajectory data indicate that incremental LOF algorithm can provide better detection of unusual trajectories in comparison to other examined techniques.

Multi-layer Background Change Detection Based on Spatiotemporal Texture Projections

In this paper we explore a multi-layer background change detection method based on projections of spatiotemporal 3D texture maps. The aim of this method is to provide a background change detection of a region viewed by multiple cameras. Camera views are projected onto a common ground plane, thus creating a spatially aligned multi-layer background. The aligned multi-layer background is subdivided into non-overlapping texture blocks, and block data is dimensionally reduced by principal component analysis. Motion detection is performed on each block, and non-moving sections of the block are clustered into multiple hyperspheres. An analysis of the clusters from spatially aligned multi-layer blocks reveal regions of changed background. This method is evaluated on surveillance videos available from PETS2006 and PETS2007 datasets.

Detecting changes in multilayered orthoimages with spatiotemporal texture blocks

A problem of detecting changes in a region viewed by multiple overlapping cameras is addressed. We are exploring a background change detection method of multilayered orthoimages using spatiotemporal texture blocks. Each camera view is projected to create a spatially aligned multilayered background orthoimages. The multilayered orthoimages are subdivided into non-overlapping blocks, and each block is represented by a 3D texture map. Texture maps are dimensionally reduced with principal component analysis. Motion detection is performed on each block and non-moving texture sections of the block are clustered into N-dimensional hyperspheres to discover changing patterns. The method evaluation is performed on publicly available surveillance video datasets.