Object, action, or scene representations that are corrupted by noise significantly impair the performance of visual recognition. Typically, partial occlusion, clutter, or excessive articulation affects only a subset of all feature dimensions and, most importantly, different dimensions are corrupted in different samples. Nevertheless, the common approach to this problem in feature selection and kernel methods is to down-weight or eliminate entire training samples or the same dimensions of all samples. Thus, valuable signal is lost, resulting in suboptimal classification. Our goal is, therefore, to adjust the contribution of individual feature dimensions when comparing any two samples and computing their similarity. Consequently, per-sample selection of informative dimensions is directly integrated into kernel computation. The interrelated problems of learning the parameters of a kernel classifier and determining the informative components of each sample are then addressed in a joint objective function. The approach can be integrated into the learning stage of any kernel-based visual recognition problem and it does not affect the computational performance in the retrieval phase. Experiments on diverse challenges of action recognition in videos and indoor scene classification show the general applicability of the approach and its ability to improve learning of visual representations.

Detecting abnormalities in video is a challenging problem since the class of all irregular objects and behaviors is infinite and thus no (or by far not enough) abnormal training samples are available. Consequently, a standard setting is to find abnormalities without actually knowing what they are because we have not been shown abnormal examples during training. However, although the training data does not define what an abnormality looks like, the main paradigm in this field is to directly search for individual abnormal local patches or image regions independent of another. To address this problem we parse video frames by establishing a set of hypotheses that jointly explain all the foreground while, at same time, trying to find normal training samples that explain the hypotheses. Consequently, we can avoid a direct detection of abnormalities. They are discovered indirectly as those hypotheses which are needed for covering the foreground without finding an explanation by normal samples for themselves. We present a probabilistic model that localizes abnormalities using statistical inference. On the challenging dataset of [15] it outperforms the state-of-the-art by 7% to achieve a frame-based abnormality classification performance of 91% and the localization performance improves by 32% to 76%.

In this paper, we propose a novel method for denoising images corrupted by the mixture of the additive white Gaussian noise and the heavy tailed noise. The proposed method is based on robust statistical approach, i.e. application of M-estimators in the combination with l1sparse regularization technique. Thus, we perform the denoising by numerical solving of the convex optimization problem. Additionally, the developed method performs denoising adaptively in order to preserve the image details, by adjusting regularization coefficients according to the local variance. The proposed de-noising scheme produces excellent results, both objectively (in terms of PSNR and MSSIM) and subjectively and outperforms state-of-the-art filters for denoising heavy tailed noise in images.

Although used extensively for object recognition, texture has for the most part been ignored as a feature used for background modelling and object segmentation. The complexity of working with texture descriptors for segmentation in videos is two-fold: the descriptive features cannot be calculated in real time and features extracted based on arbitrarily chosen regions or blocks in the frame are not stable enough to allow for building models sufficiently accurate, yet simple enough to be used for real-time segmentation. The paper proposes an approach that can be used to detect regions of texture, stable enough to be modelled using probabilistic models commonly used for foreground segmentation. Based on the evaluated stable texture regions, a discriminative texture descriptor is proposed that can be evaluated in real time. When video is captured from a stationary camera, the background is expected to be stationary to a degree and an adaptive model can be built to serve as basis for segmentation. For reasons of efficiency most adaptive models are learnt on a per pixel basis [4]. Joint domain-space modelling [3], as well as modelling of the sequence at two different scales, has been suggested as a way to enhance the segmentation and escape the limitations of single-pixel-based models[1]. The improved approaches still fail to consider the texture properties of the objects in the background and suffer from the noise introduced by neighboring objects being modelled as one. Texture of the background can be used to build models in a more informed way. Large stable regions in the texture of the background allow for larger scale modelling and more precise extraction of texture related features. Conversely, unstable regions in terms of texture correspond to transitions between different objects in the background, and the features should be extracted for each of the objects separately. An approach to background modelling and foreground segmentation derived form these principles is proposed in the paper. In each frame of the sequence stable texture regions are determined, texture descriptor (feature) values extracted using the stable regions information and used to model the background and segment the foreground using standard probabilistic approaches. The process is illustrated in Fig. 1, where the brighter patches in the stability map correspond to more stable regions, in terms of texture. Once the size of the stable-texture region at each location is determined, arbitrary descriptive features can be extracted from the frames of the sequence. In addition, one can choose any of a number of probabilistic methods [4][2] to learn the statistics of the features. The size of the stable texture regions can be determined online, using integral images [6]. Observe that:

The paper presents an efficient and reliable approach to automatic people segmentation, tracking and counting, designed for a system with an overhead mounted (zenithal) camera. Upon the initial block-wise background subtraction, k-means clustering is used to enable the segmentation of single persons in the scene. The number of people in the scene is estimated as the maximal number of clusters with acceptable inter-cluster separation. Tracking of segmented people is addressed as a problem of dynamic cluster assignment between two consecutive frames and it is solved in a greedy fashion. Systems for people counting are applied to people surveillance and management and lately within the ambient intelligence solutions. Experimental results suggest that the proposed method is able to achieve very good results in terms of counting accuracy and execution speed.

In real-world surveillance systems, where variation of light and camera parameters can sometimes severely impair the normal operation of background subtraction algorithms, better results are obtained with differencing schemes. We have earlier demonstrated that differencing of detail images produced by wavelet transformation can lead to more stable detection results. In this paper, we considerably extend that framework, by introducing the modified z-scores calculated from wavelet coefficient differences. Foreground pixels are detected as outliers in normal distribution by modified z-score test. The threshold value used in the outlier test is optimized by maximizing the precision and recall measures on several training frames. Finally, the elimination of ghosts from motion detection is done by double modified z-score testing, that is similar in idea to double frame differencing. The resulting motion detection method shows considerable resilience to changes in illumination and camera parameters and also produces a lower amount of detection errors than other motion detection methods.