Self-tuned vehicle parameters detection and estimation from vídeo sequences for traffic incident management.

• Autor(a) principal: Alessandro Moreira Fonseca
• Data de Publicação: 2005
• Tipo de documento: Dissertação
• Idioma: eng
• Título da fonte: Biblioteca Digital de Teses e Dissertações do ITA
• Texto Completo: http://www.bd.bibl.ita.br/tde_busca/arquivo.php?codArquivo=328
• topic: Processamento de imagens; Visão por computadores; Detecção; Veículos; Identificação de parâmetros; Controle de trajetória; Vigilância; Controle automático; Engenharia eletrônica

This work proposes and compares three algorithms for vehicle detection, composed of two main stages, preprocess and process. The first will proceed with the selection of the region of interest (ROI) and the background learning phase. The ROI is a narrow and transversal-to-lane rectangular area wherein detection is achieved, its dimensions provides low processing workload. The learning, named as progressive background training (PBT), is compounded of three consecutive and nonoverlapping steps that will produce the necessary parameters for the next stage. PBT is accomplished fully automatically, expect for requiring the manual specification of the three subset of training images, one for each training step. These steps relate to find, respectively, the gray-level lane limits (GLLL), the intensity edge threshold (IET) and the normalized chromaticity distortion threshold (NCDT). The first supposes that, in normal traffic conditions, most of the input images are of lane and uses this to select the first subset of sample-frames and build two intensity histograms; a statistical operator is used for the automatic threshold selection (ATS) and further lane intensity interval definition. The second step uses GLLL interval to fulfill the automatic sample selection (ASSEL) of the second input of images subset and build a bi-modal histogram of gradient magnitude wherein two unsupervised, non-parametric and non-context automatic threshold selection (ATS) methods are used to find the intensity edge threshold. One maximizes a cost function related to conditional entropy and the other maximizes the between-classes variance. The last step, NCDT, uses the former step threshold output to perform edge detection and use it as ASSEL and select the frames to model the statistics of the background and construct the histogram of the normalized chroma distortion. An ATS carries out a search into this histogram until find a predefined detection rate and find the NCDT. The second stage, named foreground detection (FD), has two main steps and uses all the thresholds and background statistical parameters computed in the previous stage. The first step relates to the generation of the ROI masks. These masks are binary and maps the state of each ROIs pixels features, as gray-level region position, edge-pixel and normalized chroma distortion. This amount of data is used in the next step of FD to perform the decision process and yield the desired vehicle detection output. A Fathy and Siyal (F&S) approach was taken as the groundwork to the three following resulting developments of this thesis: the MF&S is a modified version of F&S specially regarding to the learning phase, it is robust to lighting changes and shadow, but not to the interface of lane and shadow; the gradient magnitude aiding (GMA) approach is a special case of the previous yielding the same performance of MF&S when it is with its best tuning; and chroma and gradient magnitude aiding (CGMA) approach, fuses the former with gray-level and color information resulting on shadow and lane interface robustness. It is also shown how vehicle detection outputs can be used to compute basic traffic parameters such as vehicle count, traffic flow, occupancy and link travel time. Performance is verified by two further defined indexes, the detection and false alarm rates. Both optimal values respectively relate to the low probability to misclassify a vehicle and incorrectly classify lane as vehicle. Results show that CGMA reduce false alarms while maintaining the detection rate in comparison with GMA and MF&S algorithms.