Cognitively Inspired Real-Time Vision Core

We introduce a cognitively inspired novel binary image representation and utilize it for real-time operating computer vision core, which is capable of simultaneously detecting a specific object in an image, classifying an image region provided by an algorithm such as motion detection, and tracking multiple objects in a video. In this framework, hidden layer representations of binary receptive field neural networks are utilized to generate compact image representations for various classification based functionalities. Object detection is implemented by learning a classifier on hidden layer activities and performing sliding window based search on the image. Morever, a classification based object tracking algorithm is introduced that uses the proposed framework, whose tracking performance in standard datasets is shown to be comparable to the state of the art Email addresses: [email protected] (Ozgur Yilmaz), [email protected] (Ismail Ozsarac), [email protected] (Omer Gunay), [email protected] (Huseyin Ozkan) 1Corresponding author Preprint submitted to Journal of LTEX Templates May 7, 2015 techniques. We further propose several additional functionalities on the same core, such as sparse interest point extraction, salient motion detection, scene recognition. These set of capabilities in the arsenal, artificial vision is expected to perform the necessary fundamental operations in real time, paving a way for more complex inferences, such as geometric computations and cross-modal information fusion.