Abstract-Pedestrians are the most vulnerable participants in urban traffic. The first step toward protecting pedestrians is to reliably detect them. We present a new approach for standing-and walking-pedestrian detection, in urban traffic conditions, using grayscale stereo cameras mounted on board a vehicle. Our system uses pattern matching and motion for pedestrian detection.

Advanced driving assistance systems (ADAS) form a complex multidisciplinary research field, aimed at improving traffic efficiency and safety. A realistic analysis of the requirements and of the possibilities of the traffic environment leads to the establishment of several goals for traffic assistance, to be implemented in the near future (ADASE, INVENT, PREVENT, INTERSAFE) including: highway, rural and urban assistance, intersection management, pre-crash. While there are approaches to driving safety and efficiency that focus on the conditions exterior to the vehicle (intelligent infrastructure), it is reasonable to assume that we should expect the best results from the in-vehicle systems. Traditionally, vehicle safety is mainly defined by passive safety measures. Passive safety is achieved by a highly sophisticated design and construction of the vehicle body. The occupant cell has become a more rigid structure in order to mitigate deformations. The frontal part of vehicles has been improved as well, e.g. it incorporates specially designed "soft" areas to reduce the impact in case of a collision with a pedestrian. In the recent decades a lot of improvements have been done in this field. Similarly to the passive safety systems, primitive active safety systems, such as airbags, are only useful when the crash is actually happening, without much assessment of the situation, and sometimes they are acting against the well-being of the vehicle occupants. It has become clear that the future of the safety systems is in the realm of the artificial intelligence, systems that sense, decide and act. Sensing implies a continuous, fast and reliable estimation of the surroundings. The decision component takes into account the sensorial information and assesses the situation. For instance, a pre-crash application must decide whether the situation is of no danger, whether the crash is possible or when the crash is imminent, because depending on the situation different actions are required: warning, emergency braking or deployment of irreversible measures (internal airbags for passenger protection, or inflatable hood for pedestrian protection). While warning may be annoying, and applying the brakes potentially dangerous, deploying non-reversible safety causes permanent damage to the vehicle, and therefore the decision is not to be taken lightly. However, in a pre-crash scenario it is even more damaging if the protection systems fail to act. Therefore, it is paramount that the

Pattern matching has been extensively used for object detection and object classification in computer vision. This process tries to solve the This paper proposes a new approach for a problem of finding a match between a set of features vehicle based pedestrian detection and classification belonging to an object and a model or a set of models. system. The pedestrian detection is performed basedThese models represent the different classes that the on the 3D data by generating a density map.object can belong to. The model that best fits the set Pedestrian classification uses a pattern matching of features is used to assign a classification to the approach and exploits both 2D image information object. and 3D dense stereo information. Because 3DObject detection and object classification information accuracy does not allow the direct based on pattern matching are traditionally limited to classification of the 3D shape, a combined 3D-2D 2D image information [3]. The advantage in using the method is proposed. The 3D data is usedfor effective 2D information consists in the fact that all the generation ofpedestrian hypotheses, scale and depth information has a high degree of accuracy and a high estimation, and 2D models selection. From the 3D level of trust. This is due to the fact that the image hypothesis, the corresponding 2D image window is represents an accurate projection of the real scene selected and the 2D hypothesis is generated. The 2D(without taking into consideration the image noise hypothesis consists in the objects external edges which, in the case of quality video cameras, is not a obtained by an edge extraction and depth based significant factor). The disadvantage of using only 2D filtering process. The scaled models are matched image information is that we do not have any against the selected hypothesis using an elastic high additional spatial information about the position and speed matching based on the Chamfer distance. The the general shape of the objects concerned. A method has been tested on synthetic and real world detection system based on pattem matching using scenarios.only the intensity image will usually try to fit all the models using all the positions and all the possible scales in order to find a match in the image. This