An Accelerated Hierarchical Approach for Object Shape Extraction and Recognition

We present a novel automatic supervised object recognition algorithm based on a scale and rotation invariant Fourier descriptors algorithm. The algorithm is hierarchical in nature allowing it to capture the inherent intra-contour spatial relationships between the parent and child contours of an object by building a tree-structure of the top-level contours that make the distinctive features of the object to be recognized. A set of distance metrics are combined to measure the similarity between two objects under the hierarchical model. To test the algorithm, a diverse database of shapes is created and used to train standard classification algorithms, for shape-labeling. The implemented algorithm takes advantage of the multi-threaded architecture and GPU efficient image-processing functions present in OpenCV wherever possible, speeding up the running time and making it efficient for use in real-time applications. The technique is successfully tested on common traffic and road signs of real-world images, with excellent overall performance that is robust to low-to-moderate noise levels.