Oleksandr Lysenko

An improved algorithmofmedian flow for visual object tracking and its implementation on TI OMAP

An improved algorithm of median flow used for visual object tracking is described. The improvement consists in adaptive selection of aperture window size and number of pyramid levels at optical flow estimation. It can increase the tracking efficiency as compared to the basic algorithm, especially when dealing with small and low-contrast objects. The proposed version of the algorithm has been implemented using OpenCV library and tested on OMAP 35x EVM and BeagleBoard-xM based on Texas Instruments OMAP3530 and DM3730 processors, respectively. Analysis of improved median flow was performed over actual video sequences. The results obtained show versatility and computational robustness of the algorithm, which makes it promising for embedded application based on ARM processors.

Hardware implementation of link aggregation in networks-on-chip

The link aggregation (LAG) technique for networks-on-chip (NoC) is described and investigated in the paper. It is shown that LAG permits to improve considerably the NoC saturation threshold due to connection of neighboring routers with the aid of multiple physical links. The proposed work has three main contributions. The first is the description of a structure and principle of operation of a NoC with LAG. The second is the comparative analysis of the synthesis results for Stratix IV FPGA. It is shown that hardware costs of LAG and virtual channel (VC) routers are comparable. The third is the evaluation of average latency and saturation threshold in LAG NoC (8×8 mesh). The simulation of System Verilog models indicates that saturation threshold in proposed approach increases by 152% compared to VC NoC.

A highly efficient behavioural model of router for network-on-chip with link aggregation

A new method of link aggregation (LAG) in networks-on-chip (NoC) is investigated. This method considerably (by 152 ÷ 300%) increases a network saturation threshold by connection of topologically adjacent routers using of multiple physical links. The authors consider several ways how to improve simulation performance of digital circuits. An algorithm, which illustrates operation of high-performance SystemC model of a NoC router with LAG, is proposed. This paper also includes results of computer simulation of the suggested fast SystemC model and its synthesisable System Verilog counterpart. The results of ModelSim simulation show that usage of proposed SystemC model instead of its System Verilog analogue reduces the duration of simulation in ten times and diminish the required memory in 121 times. Herewith, the prediction error for saturation threshold of NoC does not exceed 6.1%.

Neural Networks Saturation Reduction

The saturation of particular neuron and a whole neural network is one of the reasons for problems with training effectiveness. The paper shows neural network saturation analysis, proposes a method for detection of saturated neurons and its reduction to achieve better training performance. The proposed approach has been confirmed by several experiments.

Soft Patterns Reduction for RBF Network Performance Improvement

Successful training of artificial neural networks depends primarily on used architecture and suitable algorithm that is able to train given network. During training process error for many patterns reach low level very fast while for other patterns remains on relative high level. In this case already trained patterns make impossible to adjust all trainable network parameters and overall training error is unable to achieve desired level. The paper proposes soft pattern reduction mechanism that allows to reduce impact of already trained patterns which helps in getting better results for all training patterns. Suggested approach has been confirmed by several experiments.

Improvement of RBF Training by Removing of Selected Pattern

Number of training patterns has a huge impact on artificial neural networks training process, not only because of time-consuming aspects but also on network capacities. During training process the error for the most patterns reaches low error very fast and is hold to the end of training so can be safely removed without prejudice to further training process. Skilful removal of patterns during training allow to achieve better training results decreasing both training time and training error. The paper presents some implementations of this approach for Error Correction algorithm and RBF networks. The effectiveness of proposed methods has been confirmed by several experiments.

Stripline bandpass filters with alternating coupling coefficients

This article shows the possibility in principle to design stripline comb filters with alternating signs of the coupling coefficients. The new feature of the stepped-impedance resonators of the quarterwave type was established. They provide the opposite sings of the electromagnetic coupling with adjacent resonators. Such resonators are asymmetric relative to the vertical axes, which is drawn through their middle part. The stripline passband filters with alternating signs of coupling coefficients have attenuation poles that are located to the left and to the right of the passband. Those poles improve the selectivity of such filters. The articles provide the measured characteristics of the stripline comb filter of the forth order with alternating coupling, which has the central frequency of f0 = 1835 MHz and the passband width of 90 MHz. The filter is constructed using a dielectric material with relative permittivity of εr = 92 and has the dimensions of 7.4×4.2×2 mm.

A simple way to broaden objects search area for tracking methods based on discriminative correlation filters

We propose a simple way to broaden objects search area for tracking methods based on discriminative correlation filter using two window functions of different bandwidth. Testing at the VOT-2016 Challenge confirmed that this approach provided higher robustness and accuracy of tracking, comparing to standard realizations, where only one window function was used. We created a software implementation of tracking system based on the proposed approach and estimated its performance. The proposed software realization for PC platform and modern ARM platforms can track objects in real-time.