Vasily E. Gai, Natalia A. Domnina, Roman O. Barinov, Igor V. Polyakov, Vladimir A. Golubenko, Georgy D. Kuznetsov

Nizhny Novgorod State Technical University

Abstract. The paper proposes an object detection method based on the theory of active perception. The theory of active perception provides an opportunity for implementation of pre-processing and feature description stages. The region feature descriptors consisting of spectral coefficients of U-transformation allows pne to identify regions with brightness variations. To designate a region as key, that is containing brightness differences corresponding to the object contour, a coefficient is introduced, which in combination with maximum RMS results in a unique threshold for each image. To achieve invariance to rotation and scale, the template image is subjected to scale and rotation transformations. Localization of the target object is done using the k-means method. For testing, images from the ALOI database, as well as their altered copies (multiple target objects in the image under study, images with superimposed noise) were used. For each image type, the results of the proposed method were compared with the scale-invariant feature transform method. The parameters resulting in the highest accuracy for the proposed object detection method were proposed for analysis. On noisy images, accuracy of the proposed method increased by 30% compared to the existing method. With more than one target object present on the test image, the proposed method was able to detect all target objects with an accuracy of about 96%.

image processing, object detection, theory of active perception