From Object Recognition to Object Localization

Introduction

Computer vision generally interferes with recognizing patterns and targets. A wealth of research is devoted to the building of algorithms capable of either detecting simple blob-like structures or recognizing complicated patterns. Generally, the efficiency of a pattern recognition technique depends on its ability to decode, with as much accuracy as possible, vital visual information contained in the natural environment. During the past few years, remarkable efforts were made to build new algorithms for robust object recognition in difficult environments. To this end, researchers emphasized in developing recognition paradigms based on appearance features with local potency (Nister, D., & Stewenius H. 2006, Sivic, J., & Zisserman, A. 2003). Algorithms of this field extract features with local extent that are invariant to possible illumination, viewpoint, rotation and scale changes.

Another aspect that has received much attention in the literature is to exploit the data derived during recognition with a view to provide objects’ spatial information. Apart from its identity, several other object-related characteristics, such as its distance to the camera or its pose (orientation relative to the camera’s plane), could be obtained (Thomas, A., et al. 2009, Sandhu, R., et al. 2009, Ekvall, S., 2005). As a result, assigning spatial attributes to recognized objects provides solutions to numerous technical problems. In robotics applications robots must be equipped with a sense of location and direction with a view to the efficient accomplishment of navigation or demanding pick and place tasks (Kragic, D., et al. 2005, Wong, B., & Spetsakis, M. 2000). In addition, spatial information is required in surveillance processes, where recognized targets are located in the working space of the robot. Furthermore, accurate perception of depth is mandatory in driver assistance applications (Borges, A. P. et al. 2009). Quality control procedures of industrial production frameworks demand accurate acquisition of enhanced spatial information in order to reject faulty prototypes. To sum up, the ultimate challenge for computer vision society members is the building up of advanced vision systems capable of both recognizing objects and providing their spatial information in cluttered environments.

This chapter is mainly devoted to two major and heavily investigated aspects. Initially, the current trend in recognition algorithms suitable for spatial information retrieval is presented. Several recently proposed detectors and descriptors are analytically presented along with their merits and disadvantages. Their main building blocks are examined and their performance against possible image alterations is discussed. Furthermore, the most known techniques emphasizing in the estimation of pose and location of recognized targets are presented. The remainder of the chapter is structured as follows: In Section 2, we give an overview of the current trend in object recognition techniques. A review of recently proposed pose estimation and 3D position calculation algorithms is presented in Section 3. Furthermore, at the last part of the section a comparison study of the presented pose estimation schemes is illustrated. Finally, the chapter concludes with a discussion and an outlook to the future work.