Object recognition for an autonomous wheelchair equipped with a RGB-D camera

This thesis has been carried out within a project at AR Lab (Autonomous Robot Laboratory) and IAS-Lab (Intelligent Autonomous Systems Lab) of Shanghai Jiao Tong University and University of Padua respectively. The project aims to create a system to recognize and localize multiple object classes for an autonomous wheelchair called JiaoLong, and in general for a mobile robot. The thesis had as main objective the creation of an object recognition and localization system in an indoor environment through a RGB-D sensor. The approach we followed is based on the recognition of the object by using 2D algorithm and 3D information to identify location and size of it. This will help to obtained robust performance for the recognition step and accurate estimation for the localization, thus changing the behavior of the robot in accordance with the class and the location of the object in the room. This thesis is mainly based on two aspects: • the creation of a 2D module to recognize and detect the object in a RGB image; • the creation of a 3D module to filter point cloud and estimate pose and size of the object. In this thesis we used the Bag of Features algorithm to perform the recognition of objects and a variation of the Constellation Method algorithm for the detection; 3D data are computed with several filtering algorithms which lead to a 3D analysis of the object, then are used the intrinsic information of point cloud for the pose and size estimation. We will also analyze the performance of the algorithm and propose some improvements aimed to increase the overall performance of the system besides research directions that this project could lead