Unified Optimization for Multiple Active Object Recognition Tasks with Feature Decision Tree

Visual object recognition plays an important role in the fields of computer vision and robotics. Static analysis of an image from a single viewpoint may not contain enough information to recognize an object unambiguously. Active object recognition (AOR) is aimed at collecting additional information to reduce ambiguity by purposefully adjusting the viewpoint of an observer. Existing AOR methods are oriented to a single task whose goal is to recognize an object by the minimum number of viewpoints. This paper presents a novel framework to deal with multiple AOR tasks based on feature decision tree (FDT). In the framework, in the light of the distribution of predetermined features on each object in a model base, a prior feature distribution table is firstly created as a kind of prior knowledge. Then it is utilized for the construction of FDT which describes the transition process of recognition states when different viewpoints are selected. Finally, in order to determine the next best viewpoints for the tasks with different goals, a unified optimization problem is established and solved by tree dynamic programming algorithm. In addition, the existing evaluation method of viewpoint planning (VP) efficiency is improved. According to whether the prior probability of the appearance of each object is known, the VP efficiency of different tasks is evaluated respectively. Experiments on the simulation and real environment show that the proposed framework obtains rather promising results in different AOR tasks.