Andrey Vavilin

Fast human detection based on parallelogram haar-like features

Inspired by a recent image descriptors for object detection, this paper proposed the feature description method based on set of modified Haar-like features which have parallelogram shapes. Using the proposed feature descriptors to develop a rapid detection system for human detection based on cascade structure used for boosting classifier. Specially, human detection in omnidirectional image as well as unwrap omnidirectional to panoramic image were described in this paper. The experimental results showed that the proposed method could produce high accuracy detection rate with lower false positive rate and higher recall rate than Haar-like features, and faster than HOG feature. It is efficiency with different resolutions and poses under a variety conditional such as flare illumination, clutter backgrounds, and so on.

HDR Image Generation based on Intensity Clustering and Local Feature Analysis

This paper describes a cluster-based method for combining differently exposed images in order to increase their dynamic range. Initially an image is decomposed into a set of arbitrary shaped regions. For each region we compute a utility function which is based on the amount of presented information and an entropy. This function is used to select the most appropriate exposure for each region. After the exposures are selected, a bilateral filtering is applied in order to make the interregional transitions smooth. As a result we obtain weighting coefficients for each exposure and pixel. An output image is combined from clusters of input images using weights. Each pixel of the output image is calculated as a weighted sum of exposures. The proposed method allows recovering details from overexposed and underexposed parts of image without producing additional noise. Our experiments show effectiveness of the algorithm for the high dynamic range scenes. It requires no information about shutter speed or camera parameters. This method shows robust results even if the exposure difference between input images is 2-stops or higher.

Vehicle localization using omnidirectional camera with GPS supporting in wide urban area

This paper proposes a method for long-range vehicle localization using fusion of omnidirectional camera and Global Positioning System (GPS) in wide urban environments. The main contributions are twofold: first, the positions estimated by visual sensor overcome the motion blur effects. The motion constrains of successive frames are obtained accurately under various scene structures and conditions. Second, the cumulative errors of visual odometry system are solved completely based on the fusion of local (visual odometry) and global positioning system. The visual odometry can yield the correct local position at short distance of movements but it will accumulate errors overtime, on the contrary, the GPS can yields the correct global positions but the local positions may be drifted. Moreover, the signals received from satellites are affected by multi-path and forward diffraction then the position errors increase when vehicles move in dense building regions or jump/miss in tunnels. To utilize the advantages of two sensors, the position information should be evaluated before fusion by Extended Kalman Filter (EKF) framework. This multiple sensor system can also compensate each other in the case of losing one of two. The simulation results demonstrate the accuracy of vehicle positions in long-range movements.

Fast HDR image generation from multi-exposed multiple-view LDR images

Most of the existing methods are focused on HDR image generation for static scenes. However, these methods are not effective for the scenes with moving objects. In this paper we present an algorithm that generates High Dynamic Range (HDR) images from three differently exposed Low Dynamic Range (LDR) images taken from different viewpoints. It allows generating HDR images for scenes with fast motion or from moving vehicle. Image alignment is based on optimized MTB algorithm. Error correction for occluded parts of the scene is solved by using error maps based on image correlation and bilateral smoothing weight. Fast exposure-blending technique is applied to aligned images in order to produce HDR image.

Fast HDR image generation technique based on exposure blending

In the proposed work a method for generating HDR images based on exposure blending is described. Using three differently exposed images a single image with recovered details in shadows and highlights is generated. Input images are analyzed to locate over and underexposed regions based on pixels intensity. Local contrast of input images is also considered in order to generate the output image with correct color transactions between differently exposed areas. Then images are merged using blending function. The proposed method requires a single pass thru the image to generate the result, which allows to process one image for less than 100 milliseconds. The proposed method requires no information about camera or shutting conditions, such as shutter speed or aperture size.

Recursive HDR image generation from differently exposed images based on local image properties

Dynamic range limitation of CCD-cameras may cause distortions and data loses in images. Such limitations are strongly effect to the further image processing. This paper describes method of combining information from differently exposed images for increasing dynamic range. Initially image is decomposed into set of regions. For each of region we compute detail evaluation function which represents its local properties. Namely mean intensity, intensity deviation and entropy. This function is used to detect regions with high dynamic range. The regions with high dynamic range are then recursively decomposed. This process iterates until all HDR regions are processed, or the size of these regions is too small for decomposition. During the process of decomposition we select the best exposure for each sub-region. For smoothing interregional transaction we used Gaussian-based smoothing function. Proposed technique allows recovering details in overexposed and underexposed parts of image. Our experiments show effectiveness of algorithm for the scenes with high dynamic range. Proposed method shows robust results even if the exposure difference between input images is 2-stops or higher.

Self-configuration for surveillance sensor network

This paper describes an autonomous monitoring system to supervise environmental parameters such as temperature, humidity, poisonous gases or smoke concentration, etc. This system is designed as a set of sensor nodes connected via a wireless network. The feature of this system is its ability to autonomously configure the network structure and synchronize data between nodes. This allows the network to be fault-tolerant. Each sensor node consists of three layers. The bottom layer includes a 5 to 12 volts battery and a stabilizer. The middle level consists of an 8-bit microcontroller, LCD (Liquid Crystal Display), SD (Secure Digital) card reader, and a set of sensors. The top level includes RF (Radio Frequency) communication module and GPS (Global Positioning System) module. Each sensor node is able to work as a standalone unit and as a part of a sensor network. The experimental works were performed for ensuring worth using in real time.

Automatic context analysis for image classification and retrieval

This paper describes a method for image classification and retrieval for natural and urban scenes. The proposed algorithm is based on hierarchical image contents analysis. First image is classified as urban or natural according to color and edge distribution properties. Additionally scene is classified according to its conditions: illumination, weather, season and daytime based on contrast, saturation and color properties of the image. Then image content is analyzed in order to detect specific object classes: buildings, cars, trees, sky, road etc. To do so, image recursively divided into rectangular blocks. For each block probabilities of membership in the specific class is computed. This probability computed as a distance in a feature space defined by optimal feature subset selected on the training step. Blocks which can not be assigned to any class using computed features are separated into 4 sub-blocks which analyzed recursively. Process stopped then all blocks are classified or size of block is smaller then predefined value. Training process is used to select optimal feature subset for object classification. Training set contains images with manually labeled objects of different classes. Each image additionally tagged with scene parameters (illumination, weather etc).

Contour-based algorithm for vectorization of satellite images

Process of object recognition in satellite images of high resolution is a complex task associated with a time consumption and complexity of the operators work. This paper describes an innovative approach for solving this problem. Based on monochromatic high-resolution satellite images (in the process of using data from the QuickBird satellite with a maximum resolution of 0.6 meters per pixel) geodata bitmap and vectorized output are received (shape files). The principle of object recognition in a satellite image is based on the allocation of edges in the gradient transition using a threshold filter. Obtained data is then transformed to a vector output using straight line detection and connected components analysis. The proposed method allows to process satellite images of large size with high performance. The performance of the proposed method can be improved by using GPU-based computations.

An Efficient Method for Correcting Vehicle License Plate Tilt

Tilt correction is a very crucial and inevitable task in the automatic recognition of the vehicle license plate (VLP). In this paper, according to the least square fitting with perpendicular offsets (LSFPO) the VLP region is fitted to a straight line. After the line slope is obtained, rotation angle of the VLP is estimated. Then the whole image is rotated for tilt correction in horizontal direction by this angle. Tilt correction in vertical direction by inverse affine transformation is proposed for removing shear from the LP candidates. Despite the success of VLP detection approaches in the past decades, a few of them can effectively locate license plate (LP), even when vehicle bodies and LPs have similar color. A common drawback of color-based VLP detection is the failure to detect the boundaries or border of LPs. In this paper, we propose a modified recursive labeling algorithm for solving this problem and detecting candidate regions. According to different colored LP, these candidate regions may include LP regions. Geometrical properties of the LP such as area, bounding box and aspect ratio are then used for classification. Various LP images were used with a variety of conditions to test the proposed method and results are presented to prove its effectiveness.