Vehicle Detection Using Shape And Appearance Constrained Active Basis Model

Vehicle detection as the prerequisite step of solving traffic congestion, frequent traffic accidents and other issues, has important practical significance. Due to the large variations of the vehicle itself and changes of the surrounding environment, vehicle detection has become a challenging problem. Vehicle detection is a highly active research topic in the field of computer vision for its application value.Active Basis Model(ABM) is a template matching method to detect the deformed object. It learns a template from training samples, the template consists of a certain number of Gabor wavelet elements with specified locations and orientations. As the elements are allowed to shift their locations and orientations within a specific range, ABM has a good effect in detecting deformed objects. However, vehicle is a rigid object, It appears slightly shape change. Besides, the complex surroundings such as traffic lines, non-motorized vehicles making a multiple of false negative detections. This paper not only investigates the reason of false negative detections but also introduces appearance and shape constrained active basis model,Active basis model has good effect of detecting deformed object. In order to increase the template matching score, the bases are allowed to shift in the given range without constraint in the process of detection, which leading to incompatible shape between detections and vehicle as well as false positives. As a rigid object vehicle is, detections shape and vehicles shape should be consistent. This paper introduces constraint to template elements perturbation to preserve shape, making the true positives more precise, eliminating erroneous false negatives.Active basis model uses Gabor wavelets as the base elements. Gabor wavelet is able to describe the texture and edge features of the image. Active basis model only exploits edge feature of vehicle. However, the use of edge information is insufficient to accurately describe the characteristics of vehicles in the actual complex scenes. In this paper, we analyze the appearance’s difference between the false negative detection area and the correct detection region, and propose to fuse color information into edge information, in order to constrain the element’s score in template sides, reducing the false detection rate.