Non-category-specific Object Recognition And Learning Model

Robots are entering our life as extension of modern computer technologies, whilehigh-intelligent and characteristics do be a distinct feature of personal robot.Computer vision system is one of the important structures in robot system that sensesurroundings and it is vital to acquire raw data and process for helping robot controland accomplish corresponding tasks. Note that object recognition as one ofapplications in computer vision system is greatly benefit of grasping object, sensingcircumstances, navigation and so on.It helps robot fulfill many tasks that detecting object in indoor environment. Italso an essential step for many computer vision tasks. Based on an existing method, inthis paper, a novel method by combining a new cue, the depth information, isproposed for detecting and localizing generic objects. It gives the four image cues forobjectness measures: multi-scale saliency, color contrast, edge density and superpixelsstraddling in the original algorithm which do not perform very well in indoor scene,especially for complex background and closed object. So, the depth information isintroduced as a new cue. It performs better than original algorithm without depthinformation.Basing non-category-specific object detection algorithm (objectness measures)and raw data from mono-sensor, this paper proposes a new model thatnon-category-specific object learning model. It is difference from template methodand general method. First, interesting object will be detected and located withobjectness measures. Then it should be recognize and classify by using informationexited in memory library. For learning in this paper means learning with a teacher,robot should be kept new object information that other people give so as to recognizeit at next time. This paper gives three hypotheses which are target objects have been segmented,these objects have been classified and information (e.g. ascription) of object has beenstored in the image library. According these hypotheses, three situations are given thatwhole information are stored in memory library, only category existed in memorylibrary and non information in the memory library. Then it describes model process indetail. In this paper, object is detected through objectness measures and then advancedfeatures and classify methods are used for accomplishing recognition task. At last,three experiments prove its validity.