Research And Implementation Of Object Recognition From A Single Monocular Aerial Robot

Every year natural disasters occur in some places of the world. Major disasters cause huge casualties and threaten human lives. So after a disaster, the most urgent needs have to be search and rescue of survivals. However, disaster site is very dangerous and its environment is complex, such as the collapsed house, poisonous substances, fires and other dangers. In such situations, it is not accessible or risky for humans. So rescue robots are the main rescuers. Their desired tasks are providing map data and searching for injured people. Towards this direction, this thesis uses an aerial robot whose name is AR.Drone to complete the rescue task of object recognition.This thesis presents an object recognition system based on homography (ORSH). The system is divided into two parts:obtaining image data and processing image data. The first part is done by AR.Drone. The vertical camera which is under the body of AR.Drone is taking pictures and transferring image data by wireless network while AR.Drone is flying. In the second part, the paper filters obtained images by matching color histogram and then detects features using SURF algorithm. Based on the pairs of matched features, Homography algorithm is used to recognize objects. Color histogram method computes color distributions for the whole image. So this thesis divides the whole image into4regions of the same size, and then computes color histogram for each region. The SURF features are not only invariant to image scale and rotation, but also robust to affine and illumination changes. Especially, the advantage of faster calculated performance contributes to successfully using SURF algorithm in ORSH. Experimental results show that this system achieves good results and real-time performance.This thesis also presents an improved system which is called an object recognition system based on labeling (ORSL). This system improves ORSH. ORSL employs labeling method to get contours of all objects in database and obtained images. It removes the noise of distributions on image background that computing color histogram only for labeled objects. This step contributes to getting high accuracy about the values of color histogram matching. Based on labeling results, object recognition is performed by solving the error results of feature matching. Experimental results show that ORSL gets a higher performance in general and achieves a real-time performance.