Radovan Fusek

Hierarchical blurring mean-shift

In recent years, various Mean-Shift methods were used for filtration and segmentation of images and other datasets. These methods achieve good segmentation results, but the computational speed is sometimes very low, especially for big images and some specific settings. In this paper, we propose an improved segmentation method that we call Hierarchical Blurring Mean-Shift. The method achieve significant reduction of computation time and minimal influence on segmentation quality. A comparison of our method with traditional Blurring Mean-Shift and Hierarchical Mean-Shift with respect to the quality of segmentation and computational time is demonstrated. Furthermore, we study the influence of parameter settings in various hierarchy depths on computational time and number of segments. Finally, the results promising reliable and fast image segmentation are presented.

AdaBoost for Parking Lot Occupation Detection

Parking lot occupation detection using vision systems is a very important task. Many systems use different sensors and their combinations to find out whether the parking lot or space is occupied or not. Using CCTV systems makes it possible to monitor great areas without a need of many sensors. In this paper, we present a method that uses the boosting algorithm for car detection on particular parking spaces and shifting the image to obtain a probability function of car appearance. Using the model of parking lot, we achieve occupancy of each parking space. We also experimented with the detector that is based on the histogram of oriented gradients (HOG) with a support vector machine (SVM). Nevertheless, we found some drawbacks of this detector that we describe in experiments. On the grounds of these drawbacks, we decided to use the AdaBoost based detector.

Energy-transfer features and their application in the task of face detection

In this paper, we describe a novel and interesting approach for extracting the image features. The features we propose are efficient and robust; the feature vectors of relatively small dimensions are sufficient for successful recognition. We call them the energy-transfer features. In contrast, the classical features (e.g. HOG, Haar features) that are combined with the trainable classifiers (e.g. a support vector machine, neural network) require large training sets due to their high dimensionality. The large training sets are difficult to acquire in many cases. In addition to that, the large training sets slow down the training phase. Moreover, the high dimension of feature vector also slows down the detection phase and the methods for the reduction of feature vector must be used. These shortcomings became the motivation for creating the features that are able to describe the object of interest with a relatively small number of numerical values without the use of methods for the reduction of feature vector. In this paper, we demonstrate the properties of our features in the task of face detection.

Energy transfer features combined with DCT for object detection

The basic idea behind the energy transfer features is that the appearance of objects can be described using a function of energy distribution in images. Inside the image, the energy sources are placed and the energy is transferred from the sources during a certain chosen time. The values of energy distribution function have to be reduced into a reasonable number of values. The process of reducing can be simply solved by sampling. The input image is divided into regular cells. The mean value is calculated inside each cell. The values of samples are then considered as a vector that is used as an input for the SVM classifier. We propose an improvement to this process. The discrete cosine transform coefficients are calculated inside the cells (instead of the mean values) to construct the feature vector for the face and pedestrian detectors. To reduce the number of coefficients, we use the patterns in which the coefficients are grouped into regions. In the face detector, the principal component analysis is also used to create the feature vector with a relatively small dimension. The results show that, using this approach, the objects can be efficiently encoded with a relatively short vector with the results that outperform the results of the state-of-the-art detectors.

Gradient-DCT (G-DCT) descriptors

Many feature-based object detectors have shown that the use of gradient image information can be a very efficient way to describe the appearance of objects. Especially, the gradient sizes, directions and histograms are commonly used. In this area, the histogram of oriented gradients (HOG) is considered as the state-of-the-art method. The histograms and gradient orientations are used to encode the gradient information in HOG. Nevertheless, many works have proved that the feature vector dimensionality of HOG can be reduced; particularly, the information of the gradient directions is redundant and it can be reduced. This was the motivation to encode the gradient information with the least possible redundant information. In this paper, we propose the method in which the discrete cosine transform (DCT) is used to effectively encode the gradient information; using DCT, the gradient information can be encoded with a relatively small set of DCT coefficients in which the most important gradient information is preserved. We show the properties of presented method for the case of solving the problem of face and pedestrian detection.

Hierarchical evolving mean-shift

Segmentation and filtration are widely discussed problems in image processing. Mean shift and its variants belong to the most popular methods in this area. In this paper, we propose a new variant of relatively new evolving mean shift that is based on the idea of minimization of dataset energy given by the sum of sizes of the mean-shift vectors. Our hierarchical EMS is focused on a significant reduction of computational time due to hierarchical evolution of the size of the kernel. We also present acceleration of precise point selection and vector recalculation, which can be applied also to original EMS.

Distance-Based Descriptors and Their Application in the Task of Object Detection

In this paper, we propose an efficient and interesting way how to encode the shape of the objects. A lot of state-of-the art descriptors (e.g. HOG, Haar, LBP) are based on the fact that the shape of the objects can be described by brightness differences inside the image. It means that the descriptors encode the gradient or intensity differences inside the image (i.e. edges). In the cases that the edges are very thin, the edge information can be difficult to obtain and the dimensionally of feature vector (without the method for reduction) is typically large and contains redundant information. These ills are motivation for the proposed method in that the edges need not be hit directly; the input brightness function is transformed using the appropriate image distance function. After this transformation, the values of distance function inside objects and backgrounds are different and the values can be used for description of object appearance. We demonstrate the properties of the method for the case of solving the problem of face detection using the classical sliding window technique.

Layered RC Circuit Model for Background Subtraction

The background subtraction is a technique widely used for video analysis, mainly moving object detection for surveillance systems. Such algorithms must be robust, fast and it has to be able to deal with dynamic backgrounds like water surface or moving tree branches. Also, they should be able to deal with illumination changes and objects casted shadows. Generally, in computer vision the algorithms with a physical background have the best performance. We propose an algorithm for background subtraction based on a model of layered RC circuits. We tested our method on video sequences acquired from level crossing and on commonly used datasets. Finally, we have compared the proposed method with other frequently used methods.

Layered Mean Shift Methods

Segmentation is one of the most discussed problems in image processing. Many various methods for image segmentation exist. The mean-shift method is one of them and it was widely developed in recent years and it is still being developed. In this paper, we propose a new method called Layered Mean Shift that uses multiple mean-shift segmentations with different bandwidths stacked for elimination of the over-segmentation problem and finding the most appropriate segment boundaries. This method effectively reduces the need for the use of large kernels in the mean-shift method. Therefore, it also significantly reduces the computational complexity.

Hierarchical Layered Mean Shift Methods

Many image processing tasks exist and segmentation is one of them. We are focused on the mean-shift segmentation method. Our goal is to improve its speed and reduce the over-segmentation problem that occurs with small spatial bandwidths. We propose new mean-shift method called Hierarchical Layered Mean Shift. It uses hierarchical preprocessing stage and stacking hierarchical segmentation outputs together to minimise the over-segmentation problem.