Moving object recognition by a shape-based neural fuzzy network

Moving object recognition by a shape-based neural fuzzy network is proposed in this paper. The moving objects considered in this paper include pedestrians, vehicles, motorcycle, and dogs. Given the shape of the moving object, its contour is calculated by contour following. The distance between the contour center and each contour point is calculated and smoothed. Parts of the feature vector are obtained from discrete Fourier transform coefficients of the smoothed distances. The length-to-width ratio of the object's shape, which is derived from vertical and horizontal projection of the shape of the object, is also used as a feature. Based on the feature vector, the self-constructing neural fuzzy inference network (SONFIN) is used for recognition. To verify the performance of the proposed approach, two experiments were performed. In the first experiment, the shape of an object was extracted manually. In the second experiment, the shape of an object was extracted automatically from a series of image processes, including gray-based and edge-based image subtractions and morphological operations. The experiments show that the proposed approach can recognize moving objects with high accuracy. SONFIN performance is also shown to be better than back-propagation neural network and radial basis function network performance.