Perception is a sensory motor process and, as such, parallel processes are supposed to manage visual focus and action selection. In order to model the perception process, this paper proposes and explores a formalism derived from the behavior of ants: the self organization of collective intelligence displayed by social insects.

The ant-foraging metaphor is the basis of the control algorithm that drives virtual agents that collectively exhibit an effective attention policy. The motivation of applying this metaphor to visual processes is described in Sections 1 and 2, with reference to literature in a broad area. However, the paper does not use real images as input to the attention process; it continues the tradition of virtual agents and artificial life. The input images considered are simple body-centric two-dimensional images of the local environment.

The method models a swarm of simple homogeneous agents (p-ants), probabilistically created to search the visual field. They interact through stigmergy for better coverage. The perceptual process and action selection process communicate for several iterations, until a motor action is reached.

As explained in the next section, the action selection process sends a message to the perceptual one regarding the utility of performing each possible action. The utility is computed from the direction of motion and the free space in the local environment. The perceptual process uses the utilities to search for obstacles. Obstacles are used to compute the radial sectors that contain free space, and this angle range is sent back to the action selection process. The iterations continue for a given maximum number or until the action with the highest utility stabilizes. This action is passed to the low-level motion controller.

The perceptual process combines different behaviors to move one step in the preferred search direction, to track an obstacle, or to walk around an anchor point. The action selection process defines the direction to take using a radial pattern, depending on empirically defined scale factors.

The experiment to validate the model is conducted on a simulated wheeled mobile robot. The world is binary, with black and white pixels identifying obstacles and free space. Two environments are discussed: one has large obstacles and the other has many small obstacles. The results are good with the same parameterization and surpass other similar approaches. The discussion remarks how the p-ants formalism is more apt than other methods to produce self-organization in neural structures.

No simulation of the real image acquisition method is done; the authors intend to extend the model to convert a three-dimensional model into two-dimensional images and to use pixels of different colors. Although these extensions can be easy in a simple world, I am guessing that the real structure and statistics of images can be more difficult to capture than considered here. Moreover, the literature about the neural organization of animal or human visual systems proposes methods that make the architectural organization of the neurons the central issue. The bridge between the architecture-based method and the self-organization-based method is still far off.