Vehicle Segmentation Using Evidential Reasoning - Intelligent Robots and Systems, 1997. IROS '97., Proceedings of the 1997 IEEE/RSJ International Co

This paper proposes a segmentation algorithm by means of an evidential reasoning to segment moving vehicles in front of the moving our car in a road traffic scene. Generally, an evidential reasoning finds the perceptually known evidences of a target and updates a probabilistic expectation for the target to be in an image. Since a noise image produces unreliable features and degrades the detection and localization, selecting image primitives which are less sensitive to noise and well represent the evidences is important. We carry out this task by the probabilistic integration of image features based on rnaxirnuni a posteriori(MAP) probability that combines the prior and likelihood probabilities using Bayes’ rule. 1: Introduction In this paper, we are interested in monocular graylevel visual sensing for segmenting moving vehicles in front of the moving our car in a road traffic scene. The segmentation is carried out by evidential reasoning based on the probabilistic integration of low-level image features. Recently, many researchers have been working on the analysis of the road traffic scene since the late 1980’s in accordance with the increasing interest about road traffic safety and intelligent vehicle development [ 1,2,4]. However, most previous researches have shown their feasibility in very limited environments and provided poor estimates due to noisy sources resulting from variable illumination, dynamic state, the diversity and complexity of scenes [ l , 2, 3 , 41. What is worse, since the viewer moves in a dynamic road scene, it is difficult to extract only the regions corresponding to moving vehicles using familiar methods of motion segmentation[2]. These problems have made the visual perception of outdoor road environment difficult and challenging topic of computer vision. Evidential reasoning is a method for combining information from different sources of evidence to update probabilistic expectations. Combining of evidences thus can be quantitatively described by combining of probabilistic expectations of many sources of information, including a color, a texture, a shape and a prior knowledge in a flexible way to achieve recognition. Xie et al. [4] have applied a perceptual organization to extract road vehicles in a dynamic traffic environment. They considered the evidence for a road vehicle as a cluster of nearly vertical and nearly horizontal line segments, which make neighbors each other. This evidence, however, is not well consistent with various kinds of roads. Furthermore, the vertical line segments are rarely extracted on a vehicle. It means that the selection of evidences of a target is more important than the application of evidential reasoning. Using some appearance properties of the object of interest in an image, we determine evidences and search for features which satisfy the evidences. While this method provides good results in noisy environments, it is domain specific. We determine the evidences for a leading vehicle within a driver’s view such as: first, there is a sharp intensity change at the boundary of a vehicle; second, a vehicle has symmetric property. For the segmentation, we take into account the following reasonable assumptions: I) A vehicle lies on a traffic road. 11) Optical axis of a CCD camera is parallel to the ground plane of a road. 111) The intensity distribution with respect to road and sky is sufficiently homogeneous between two successive images. IV) The intensity distribution between a road and a vehicle is not similar. We also limit the problem by considering the following two practical constraints: I) The entire shape of the leading vehicle in front of a viewer is contained in an image. 11) The road should be seen in an image. Proc. IROS 97 0-7803-41 19-8/97/$1001997 IEEE