Marcus V. Lamar

Neural network for travel demand forecast using GIS and remote sensing

Describes an application of neural networks in the development of a travel forecast model for transportation planning. The model intends to quantify trips within the urban area through the representation of the land use-transportation system interaction. The data to express such a complex interaction is mainly obtained from remote sensing images that are processed in a geographical information system. We present the models basic formulation and the results of a case study conducted in the Boston metropolitan area.

A new approach to multiple vehicle tracking in intersections using harris corners and adaptive background subtraction

This work presents a new methodology based on video processing for recording and counting vehicles in intersections and urban roads, allowing the control and management of urban traffic and support to inspection of public road security in large cities. The aim of this work is to show the viability of computer vision techniques for the construction of a portable equipment able to perform urban vehicles flow counting and classification into routes in an automatic way and in real time. Image processing techniques, such as background subtraction, definition and tracking of object features are used in the prototype development and implementation. The obtained results show the viability of use of the proposed portable monitoring system in real-time, achieving correct rates in the vehicle counting of about 96% for simple roads and 72% at complex intersections.

Real-time hand postures recognition using low computational complexity Artificial Neural Networks and Support Vector Machines

This paper proposes two main techniques for reduce computational complexity on artificial neural networks, using piecewise linear activation function, and support vector machines built on a probability based binary tree. These methods are compared with well-known classifiers based on the computational complexity, correct rate and time taken to process the required information. The results show that probability based binary tree SVM has an equivalent recognition rate and is faster than ANNs.

Dynamic Visualization of Transit Information Using Genetic Algorithms for Path Schematization

In this paper, we present a genetic algorithm for path octilinear simplification. The octilinear layout, recognized worldwide in metro maps, has the special property that edge orientations are restricted to eight angles. The proposed search technique combines possible solutions to find a solution with a desired balance between faithfulness to the original shape and reduction of bends along the path. We also aim the genetic algorithm to real-time response for dynamic web visualizations so we can experiment on how algorithms for the visualization of schematic maps can be availed in a context of mobile web devices in order to empower efficiency in transmitting transit and navigation information. A prototype of a web application and real transit data of the city of Castellon in Spain were used to test the methodology. The results have shown that real-time schematizations open possibilities concerning usability that add extra value to schematic transit maps. Additionally, performance tests show that the proposed genetic algorithms, if combined with topological data and scale variation transformation, are adequate to sketch bus transit maps automatically in terms of efficiency.

Design and Optimization of Digital Circuits by Artificial Evolution Using Hybrid Multi Chromosome Cartesian Genetic Programming

Traditional digital circuit design techniques are based, for the most part, on top-down methods, which use a set of rules and restrictions to assist the construction of the project. Genetic algorithms, on the other hand, haven proven themselves to be a very useful tool for solving high complexity problems, relying on a bottom-up methodology. This paper proposes a new design algorithm, named HMC-CGP, which operates by first finding a functional solution by using the MC-CGP method. Then, optimizes it by using standard CGP approach. Test circuits used include 1 and 2 bit full adders, 2 bit multiplier and 7 segment hexadecimal decoder. Obtained results show that by making use of faster convergence granted by the MC-CGP mechanism together with an optimization strategy generates novel approaches for those circuits, with results showing a logic gate and transistor usage reduction of upi¾?to 60.8i¾?%.

Uma Comparação Entre Redes Neurais Artificiais E Máquinas De Vetores De Suporte Para Reconhecimento De Posturas Manuais Em Tempo-Real

This paper presents a comparison between different artificial intelligence techniques used for finger spelling recognition on a real-time system, with emphasis on Support Vector Machines and Artificial Neural Networks but also using other wellknown methods. The comparison is based on the computational complexity, correct rate and time taken to process the required information. The tests are based in two models of ANN and a binary tree structured multi-class SVM. The results show that SVM has an equivalent recognition rate as ANN and it is computationally cheaper and faster than ANN.

A study of the ICP algorithm for recognition of the hand alphabet

Due to high demands of accuracy and efficiency while dealing with lots of complex and ambiguous gestures, the sign language recognition is still a hard problem to vision-based systems. From depth images acquired by a recent RGB-D sensor, this paper proposes a novel methodology for the recognition of the American Sign Language hand alphabet, which is underlain by a detailed analysis of the Iterative Closest Point (ICP) algorithm, applied here as a tridimensional shape matching procedure. The evaluation of the inputs and outputs of the ICP technique combined with the interpretation of the conducted experiments contribute significantly for advancing the state of the art, by identifying in which conditions the alignment can be used in compliance with the requirements of the pattern recognition.