We present a system for object recognition that is largely inspired by physiologically identified processing streams in the visual cortex, specifically in the ventral stream. It consists of neural units organized in a hierarchy of layers with encoding features of increasing complexity. A key feature of the system is that the neural units learn their preferred patterns from visual input alone. Through this "soft wiring" of neural units the system becomes tuned for target object classes through passive visual experience and no labels are required in this stage. Object labels are only introduced in the last step to train a classifier on the system's output. While this tuning process is purely feed-forward we also present a neural mechanism for back projection of the learned image patterns down the hierarchical layers, and demonstrate how this feedback can be used to stabilize the system in the presence of noise. We test the neural system with natural images from publicly available data-sets of natural scenes and handwritten digits.

Distance estimation of visually simulated self-motion is difficult, because one has to know or make assumptions about scene layout to judge ego speed. Discrimination of the travel distances of two sequentially simulated self-motions in the same scene can be performed quite accurately (Bremmer and Lappe 1999; Frenz et al., 2003). However, the indication of the perceived distance of a single movement in terms of a spatial interval results in a depth scaling error: Intervals are correlated with the true travel distance, but underestimate travel distance by about 25% (Frenz and Lappe, 2005). Here we investigated whether the inclusion of further depth cues (disparity/motion parallax/figural cues) in the virtual environment allows more veridical interval adjustment. Experiments were conducted on a large single projection screen and in a fully immersive computer-animated virtual environment (CAVE). Forward movements in simple virtual environments were simulated with distances between 1.5 and 13 m with varying speeds. Subjects indicated the perceived distance of each movement in terms of a depth interval on the virtual ground plane. We found good correlation between simulated and indicated distances, indicative of an internal representation of the perceived distance. The slopes of the fitted regression lines revealed an underestimation of distance by about 25% under all conditions. We conclude that estimation of travel distance from optic flow is subject to scaling when compared to static intervals in the environment, irrespective of additional depth cues.

An algorithm for online distance computation based on images taken in a sewer by a robot-inspector is presented. Modern concrete sewers are made of standard cylindrical pipes separated by manhole junction areas. Circular structures typically originate at the pipe ends and entrances as well as at the joints of the pipe sections. These structures provide regular marks on the sewer images and can be used for their 3D interpretation. After each robot motion, we extract these circular structures from camera images, and recover their distance relative to the robot. The use of a laser-based orientation method to ensure almost parallel viewing direction of the robot head with respect to the pipe axis drastically simplifies the necessary image processing operations. The method is applied to real image sequences taken in a sewer test-net. Our results show that pipe ends and junctions can be reliably detected by the robot. We intend to use the approach to facilitate robot navigation for sewer inspection.

A practical algorithm for 3D interpretation of images taken in a sewer by a robot-inspector is presented. Modern sewers made of concrete present regular circular structures originating at the pipe ends as well as at the joints of the pipe sections. These pipe ends and joints provide regular marks on the sewer images and can be used for their 3D interpretation. The key idea of the paper is to extract from their image projections the 3D orientations of the circular structures and their distances relative to the robot. We adopt for this purpose a formula for the possible position of a calibrated camera on the basis of the image of a circle of known size. We describe a sequence of image processing operations that perform circle extraction, robust conic fitting and distance calculation based on original images. The reported results are preliminary in the framework of the project aimed at developing an autonomous robot for sewer inspections. We intend to use the approach to facilitate the orientation of the robot while moving in the sewer.