Vision based automatic traffic condition interpretation Filipe Alves*, Manuel Ferreira*, Cristina Santos* *Department of Industrial Electronics University of Minho Campus de Azurem – Guimarães, Portugal Telf: +351253510190; fax: +351253510189; email: nalves@dei.uminho.pt; mjf@dei.uminho.pt, cristina@dei.uminho.pt Abstract- Traffic flow, analysis and control is gaining high relevance, as the number of circulating vehicles continuously increases. This article proposes a computer vision based platform, which automatically detects vehicles in order to infer the traffic conditions. The developed real time detection algorithm is based on a dual background subtraction technique, incorporating the one known has Codebook and an edges one. These two layers interact mutually allowing the compensation of individuality weaknesses. The traffic flow parameters are extracted comparing the detected vehicles with a known model of the road lanes, which can be automatically generated based on the vehicles trajectory analysis over time. The achieved results demonstrate that the developed algorithm is able to correctly understand the traffic flow state, even in the presence of adverse situations that are typical of an outdoor application. I. INTRODUCTION Transportation ways play a fundamental role in today country’s development. With traffic roads being one of the most valuable ways, optimized traffic management and control becomes vital, especially at road intersections. To assist in traffic flow control, we can find the semaphores. These devices are usually pre-programmed or commonly assisted by inductive loops. This work proposes a low cost computer vision approach, capable of detecting the presence and flow of vehicles in order to infer the traffic conditions. Several advantages can be pointed out when compared to the inductive loop approach, namely the much higher detection area, the lower installation and maintenance costs, and the higher functionality capability. Vehicle detection is part of a broader class named visual surveillance. Usually, the objects we are trying to detect in visual surveillance (cars and persons) share the common feature of being in movement. A temporal analysis is therefore suitable for this task. Among the techniques that use it, background subtraction (BS) is one of the most studied and implemented. In its core there is a model of the field of view without the objects of interest (IO), named background (B), from which the actual image is compared with, resulting in the foreground (F) mask containing the IO. Because of its temporal analysis and scene recreated model, these techniques suffer from a number of problems characterized by changing the scene appearance. These problems, well described in [1], are: moved objects, time of day, light switch, waving trees, camouflage, bootstrapping, foreground aperture, sleeping person, waking person and shadows. Several background subtraction algorithms have been proposed throw out the time [2][3][4][5][6][7][8][9]. Special attention is given to: MOG [10], Wallflower [1] and Codebook [11]. In 1999 Chris Stauffer and W.E.L. Grimson presented the mixture of Gaussians (MOG) model, which would turn out to be the most widely and referenced background subtraction technique. On this pixel by pixel model, each pixel is represented by a group of K gaussians according to its recent observed values. At each new image, the actual pixels value is compared against his associated gaussian distributions. If it stays more than γ standard deviations of a gaussian it is labeled has foreground. Every gaussian has a weight indexed. The background model is considered to be formed by the minimum gaussians capable of explaining T percentage of the data. The Wallflower presents some of the best theoretical aspects about background subtraction. The principals, the problems and associated implications and the reference to the “need” of interconnection between several perception levels are Wallflower main contributions. The algorithm implementation has in the pixel level, a wiener filter designed to predict the pixel’s value in the next image, adapting this way, to non-static behaviors. Two more layers, one at the region level, and the other at the image level compose the final algorithm. More recently, Davis et al., proposed the Codebook (CB) model. A representation based on a quantification/clustering technique models the background not on a frequency of occurrences basis, but on a quasi-periodicity reappearance observations. Codebook is part of our traffic detection platform, thus, it will be further described. II. IMPORTANT DECISIONS PRIOR TO BACKGROUND SUBTRACTION Prior to background subtraction, some considerations with direct impact in background subtraction output performance need to be analyzed. 978-1-4244-7300-7/10/$26.00 ｩ2010 IEEE 549