Artificial visual attention has been an active research area for over two decades. Especially, the concept of saliency has been implemented in many different ways. Early approaches aimed at closely modeling saliency processing with concepts from biological attention to provide (at least in the long run) general-purpose attention for technical systems. More recent approaches have departed from this agenda, turning to more specific attention-guided tasks, most notably the accurate extraction of salient objects, for which large-scale ground truth datasets make it possible to quantify progress. While the first type of models is troubled by weak performance in these specific tasks, the second type, as we show with a new benchmark, has lost the ability to predict saliency in the original sense, which may be an important factor for future general-purpose attention systems. Here, we describe a new approach using growing neural gas to obtain pre-attentional structures for a scene at an early processing stage. On this basis, traditional saliency concepts can be applied while at the same time they can be linked to mechanisms that make models successful in salient object detection. The model shows high performance at predicting traditional saliency and makes substantial progress toward salient object detection, although it cannot reach the top-level performance of some specialized methods. We discuss the important implications of our findings.

In recent years, place recognition techniques have been extensively studied in the domain of robotic mapping, referred to as appearance-based mapping. Nonetheless, the majority of these methods focus the challenges of place recognition in offline or supervised scenarios, which in certain conditions, e.g., unknown environments, is infeasible. In this paper, we address the challenges of online place recognition and demonstrate the general applicability of our approach in versatile environments. To this end, a modified growing self-organizing network of neurons is proposed, which incrementally adapts itself to learn the topology of the perceptual space formed by gist features. Given a query image and the network state at any time instant, the expected activity of the network is estimated using a proposed Bayesian framework, and the current place is categorized as familiar or novel. Exhaustive experiments on 11 challenging sequences signify the strength of our algorithm for a reliable and real-time place recognition on routes as large as 18 km. Compared to many state-of-the-art approaches, our method does not need offline training or environment-specific parameter tuning.

This paper presents a new concept, the Admissible Gap, for collision avoidance. An admissible gap AG is defined as the gap that a robot may safely pass through, while obeying the shape and motion constraints. By employing this concept, a new obstacle avoidance approach was developed, improving the navigation performance in unknown cluttered environments. Unlike most state-of-the-art methods, our approach explicitly considers the robot shape and kinematic constraints rather than adapting a method originally designed for a holonomic pointlike robot. Experimental results demonstrated the power of the proposed AG approach. Moreover, a comparison with state-of-the-art methods showed that the AG approach generates more efficient, safer, and smoother trajectories.

Recent research in appearance-based mapping has introduced a diverse range of techniques to deal with environments under varying conditions. Common to almost all the existing methods is that they expect a supervised or offline training. Furthermore, some approaches match sequences, assuming constant velocity over the routes. This work addresses the challenges of appearance-based mapping in an online setup without making assumptions or acquiring prior knowledge of the environment. For this purpose, we exploit the topology preserving capability of self-organizing neural networks and learn the perceptual representation of environments using GIST features. Due to the fact that real-world environments are complex and large-scale, we let the network grow while accounting for the amount of error in the network due to perceptual differences among the places. Given the current state of the network and a query image at any time instant, it is possible to identify whether a place comes from the visited location by computing maximum a posteriori estimate over the network. The extensive experiments on three standard datasets, St. Lucia (4 videos), KITTI (2 sequences), and the Oxford City Center dataset, demonstrate the strength of our approach for real-time place recognition while concurrently learning the spatial representation of scenes.

Perception is the main key in enabling robots to react to and interact with their environment. Particularly, for multi-floor operations, the robot must robustly detect and localize stairs to allow for safe climbing. In this paper, we develop a graph-based stairway detection method for point cloud data, that can detect a large variety of stairways. Our approach first segments planar regions and extracts the stair tread- and stair riser-shaped segments. With these segments, a dynamic graph model is initialized that is used to detect stairs including the railing system in the surroundings. We show that our system can accurately detect and localize different stairways from a variety of different positions, including descending stairs. Our system's accuracy is higher than those of most state-of-the-art stairway detection methods even in case of sparse point cloud data.

Most of the variational optical flow methods are based on the well-known brightness constancy assumption or high-order constancy assumptions to implement the data term in the optimization energy function. Unfortunately, any variation in the lighting within the scene violates the brightness constancy constraint; in turn, the gradient constancy assumption does not work properly with large illumination changes. This paper proposes an illumination-robust constancy based on a robust texture descriptor rather than the brightness constancy. Thus, the similarity function used as a data term was obtained from extracting texture features through the local directional pattern descriptor for two consecutive frames within the duality total variational optical flow algorithm. In addition, a weighted nonlocal term that depends on both the color similarity and the occlusion state of pixels is integrated during the optimization process to increase the accuracy of the resulting flow field. The experimental results show a qualitative comparison with the proposed approach and yield state-of-the-art results on the KITTI, Midleburry, and MPI-sintel data sets.

A new reactive collision avoidance approach for mobile robots moving in cluttered and complex environments was developed and implemented. The novelty of this approach lies in the creation of a new method for analyzing openings in front of the robot that highly reduces their number when compared with the Nearness-Diagram Navigation (ND) technique, particularly in complex scenarios. Moreover, the angular width of the chosen (selected) gap with respect to the robot vision is taken into consideration. Consequently, oscillations are alleviated, the computational complexity is reduced and a smoother behavior will be achieved. Our technique adjusts the motion law proposed in the Smooth Nearness-Diagram Navigation (SND) method to generate safer paths for the robot by considering the ratio of threats on its sides and applying stricter deviation against an obstacle as it gets closer to the robot. Hence, the problem of deadlock occurring in narrow corridors, with high threats on one side and low threats on the other, is solved without affecting the smoothness behavior. Simulation and experimental results demonstrate the power of the proposed approach.

Visual attention is one of the important phenomena in biological vision which can be followed to achieve more efficiency, intelligence, and robustness in artificial vision systems. This paper investigates a region-based approach that performs pixel clustering prior to the processes of attention in contrast to late clustering as done by contemporary methods. The foundation steps of feature map construction for the region-based attention model are proposed here. The color contrast map is generated based upon the extended findings from the color theory, the symmetry map is constructed using a novel scanning-based method, and a new algorithm is proposed to compute a size contrast map as a formal feature channel. Eccentricity and orientation are computed using the moments of obtained regions and then saliency is evaluated using the rarity criteria. The efficient design of the proposed algorithms allows incorporating five feature channels while maintaining a processing rate of multiple frames per second. Another salient advantage over the existing techniques is the reusability of the salient regions in the high-level machine vision procedures due to preservation of their shapes and precise locations. The results indicate that the proposed model has the potential to efficiently integrate the phenomenon of attention into the main stream of machine vision and systems with restricted computing resources such as mobile robots can benefit from its advantages.

This paper presents a new approach for providing visual attention on robot vision systems. Compared to other approaches our method is very fast as it processes regions rather than individual pixels. The proposed method first builds a list of regions by applying a shade and shadow tolerant segmentation step. The features of these regions are computed using their convex hulls in order to simplify and accelerate the processing. Feature values are stored within the records of respective regions instead of constructing a master map of attention. Then an algorithmic method is applied for finding the focus of attention in contrast to mathematical approaches used by existing models. Experiments conducted on simulated and real image data have not only demonstrated the validity of the proposed approach but have also led to the establishment of a comprehensive robotic vision system.

This paper presents a new stereo algorithm for computing dense disparity maps from stereo image pairs by a global cost relaxation, realized as an optimization problem, where the disparity map is the momentary state of a dynamic process. Following the natural role model of the human visual system, we assign a set of possible disparities to each image pixel described by cooperating probability variables. In the first step a correlation-based similarity measure is performed to initialize the relaxation process. The relaxation itself is formulated as an optimization of a global cost function taking into account both the stereoscopic continuity constraint and considerations of the pixel similarity. A special formulation guarantees the existence of a unique cost minimum which can be easily and rapidly found by standard numerical procedures. In a post-processing step, occluded areas are detected and a sub-pixel precise disparity map is computed.

We present a rapid-prototyping environment for functional verification and test of digital signal processing algorithms. The environment consists of a Virtex-ll device on a PCI-card and an appropriate generic software backend which is used to pre- and post-process the data and to transfer it to the FPGA and pull the results from it. It is designed to meet real-time requirements by means of interleaving block-transfers to and from a large on-board memory. We use the system for the development and test of audio signal processing applications. The implementation and test of the gammatone-resynthesis algorithm is described as an exemplary algorithm that has been tested within the environment. The presented system is part of a software framework for rapid development of power optimized audio signal processing applications on behavioral level using library elements.

Mobile robot platforms and sensor mountings are susceptible to accidental contact with undetected objects and to degradation of sensor alignment after extended periods of mechanical operation. For correct sensor data interpretation and to accomplish sensor fusion, this has to be taken into account. Most sensor calibration techniques are computationally expensive or cannot be performed unattended, and therefore are not well suited for autonomous robots. However, by employing additional knowledge about the robot's construction (and sources of error), sufficient accuracy can be maintained by using a simpler adjustment procedure. This paper contributes an enhanced version of the preprocessing algorithm presented by Iocchi (1999) for automated computation of camera radial distortion with fewer demands on the image, as well as a method for self-controlled sensor adjustment by improving the matching of such preprocessed multisensor data.

Models of visual attention provide a general approach to control the activities of active vision systems. We introduce a new model of attentional control that differs in important aspects from conventional ones. We divide the selection into two stages, which is more suitable for the system as well as explaining different phenomena found in natural visual attention, such as the dispute between early and late selection. The proposed model is especially designed for use in dynamic scenes. Our approach aims at modeling as much of a general active vision system as possible and designing clean interfaces for the integration of the remaining specific aspects needed in order to solve specific problems.

This paper presents a flexible communication module for low-level as well as high-level image processing operations. It allows a good separation of data communication and data processing and thereby reduces the necessary amount of work for the implementation of parallel image processing algorithms. It supports heterogenous processor systems. It has been successfully used for the parallel implementation of a hierarchical image transition and for its symbolic analysis on a 9-node transputer image processing system. Experimental results in the field of traffic sign detection are discussed.

Objects arbitrarily positioned within the working space of our MANUTEC r2 robot are detected and foveated by a colour TV-camera mounted close to the robot's hand. Previously learnt objects are recognized and gripped while unknown objects are included into the system by feeding the name of the object. Neural representations of regions, contours, corners, and colours are learnt by only one presentation. Invariances are due to parametric mappings controlled by estimated values of orientation and distance. All components of the architecture are massively parallel, and the system provides high reliability and accuracy.