Mobile Robots Target Detection And Localization Based On RGB-D

Visual information is vital for mobile robots to perceive environment and move autonomously. This paper addresses robotic vision system, multi-views target detection and target localization.The vision system is divided into hardware-and software architecture. In the aspect of hardware, a Pioneer 3-AT mobile robot is chosen as the research platform. A PTU-46 device equipped on the robot can make the camera rotate to all directions to observe the entire environment. In terms of software architecture, the rapid development of robotics makes the need of code reuse and modularity become more and more intensive. The software system of mobile robot is based on ROS using node structure, which allows a function to be divided into a number of small features and to be combined freely. The program can thus be designed independently and combined in real time.Targets appearing in the horizon of mobile robot may have different postures. In order to let the robot to detect different postures of the target, the Adaboost algorithm is used to construct a classifier based on Harr-like features firstly. Secondly, the classifier based on the viewpoint is trained against the problem of multi-view target detection. Finally, multi-view target detection is implemented through the combination of various sub-classifiers. Experimental result shows effectiveness of this method.The accurate target location is difficult to achieve using 2-D images. However, a precise localization can be given based on the depth information of the object. For target localization problem, in this thesis limitations of the pinhole model in the positioning problem is analyzed firstly. Secondly, the significance of image’s depth and its data structure are analyzed. Thirdly, the 3D coordinate of the pixel is deduced according to the distance information. Finally, target localization is achieved. Experimental result shows feasibility of this method.