Active object recognition using vocabulary trees

Abstract

For mobile robots to perform certain tasks in human environments, fast and accurate object classification is essential. Actively exploring objects by changing viewpoints promises an increase in the accuracy of object classification. This paper presents an efficient feature-based active vision system for the recognition and verification of objects that are occluded, appear in cluttered scenes and may be visually similar to other objects present. This system is designed using a selector-observer framework where the selector is responsible for the automatic selection of the next best viewpoint and a Bayesian 'observer' updates the belief hypothesis and provides feedback. A new method for automatically selecting the 'next best viewpoint' is presented using vocabulary trees. It is used to calculate a weighting for each feature based on its perceived uniqueness, allowing the system to select the viewpoint with the greatest number of 'unique' features. The process is sped-up as new images are only captured at the 'next best viewpoint' and processed when the belief hypothesis of an object is below some pre-defined threshold. The system also provides a certainty measure for the objects identity. This system out performs randomly selecting a viewpoint as it processes far fewer viewpoints to recognise and verify objects in a scene.

Probabilistic object and viewpoint models for active object recognition

“…A wide range of general frameworks for active vision and active sensing have been explored, including information theoretic and Bayesian approaches [3] [6] [4] [5] [12], discriminative approaches [13] [14], and approaches based on other theoretical models such as possibilistic and DempsterShafner theory [15], [16]. We adopt a Bayesian framework due to its flexibility in incorporating diverse modeling choices in a principled manner.…”

“…Our framework is most directly related to that of [12]. This method is also based on SIFT and, drawing on the techniques of [10] [7] for non-active recognition, incorporates geometric structure by filtering the features processed at a given view using the Hough transform to identify the most likely transformation from a training example.…”

“…However, the method does not explicitly include the transformation as part of the probabilistic model, and further does not model the background or occlusion process. We deal with these issues by introducing a background distribution and latent occlusion and transformation variables, and incorporate a distribution over both object and pose, unlike [12].…”

“…Following [12], we also make use of a vocabulary tree representation [11] for the low-level features in our models. This has shown to be effective in both 2D and 3D recognition tasks [11], and also in Simultaneous Localization and Mapping (SLAM) approaches for matching similar images and for loop closure [17].…”

“…= 1 for a true condition and 0 otherwise (the Iverson bracket), and p a and p b are parameters of the distribution controlling the probabilities when that node n is not seen and is seen respectively (which relate to the p o and p no parameters in [12]). The likelihood for a test image with features f ind is then formed simply by treating all observed visual words as independent draws from Eq.…”