Stereo Vision-Based Object Matching, Detection, and Tracking: A Review

Introduction

Computer vision is a subject, which attempts to recreate the human vision by fabricating models which seem to have related properties to visual observation. It also tried to develop models which seem to have related characteristics to graphical perceptions. Stereo vision is an essential component of computer vision. It is the extraction of 3D information from digital images by examining the relative positions of the objects captured by the dual cameras. With the help of stereo vision, it is possible to reconstruct, either partially or fully, a 3D scene from two or more images that have been captured under marginally dissimilar angles. There are two main categories of computer vision: plane and stereo vision. The most notable difference between them is the depth information i.e. distance of the objects from the cameras. This is not possible to be detected by a single camera as it requires the use of dual cameras. Each lens captures its own view and then two independent images are sent to the system for processing. The system compares the images while shifting the two images together over top of each other to find the parts that match. The shifted amount is called disparity. The disparity at which objects in the image best match is used by the system to calculate their distance. When both the images are processed by the system, they are combined into a single image by matching up the similarities and then adding in the minor differences. These minor differences between the captured images combine to give out a relatively bigger difference in the resultant image. This combined image is greater than the aggregate of its parts. It is a 3D stereo image.

Innovative work in the field of Computer Vision began in the early 60’s when Robert from MIT had successfully finalized a 3D scene analysis project(Weiss, 1999). In his project, in order to achieve the 3D scene analysis, 2D image processing had been employed. This particular project had long been considered to be the origin of the stereo vision technique. Nowadays, a comprehensive stereo vision system could be created by concealing the initial steps from capturing images to the final step of recreating the visual surface of the objects. The concept of ‘computation stereo’ was first proposed by Barnard and Fisher(Arnaud, 2004) who explained that it covered the topics of image matching, depth information, image acquisition and also feature extraction. This has led to the conclusion that a stereo vision system with dual cameras can be utilized at the same time to capture the left & right images effectively so as to obtain the required depth information to be used in a range of applications such as a video surveillance system, mimicking the human vision for the visually impaired, for robotics, to control unmanned vehicles, for security purposes, virtual reality and 3D TV etc.

In a stereo vision process, the primitives extracted from the images that are being matched such as segments, pixels, regions etc. is found to be the most significant phase. There are two extensive types of matching methods (Banks, Bennamoun, Kubik, & Corke, 1997) .The first one makes use of pixel neighborhood correlation method, which generates a dense disparity map, whereas the other method makes use of matching based on characteristics, in this case, generating a disparity map that is comparatively sparse. In any case, the stereo matching based on edge points utilizing linear images is given more importance.