One of the main challenges in robotics is navigating autonomously through large, unknown, and unstructured environments. Simultaneous localization and mapping (SLAM) is currently regarded as a viable solution for this problem. As the traditional metric approach to SLAM is experiencing computational difficulties when exploring large areas, increasing attention is being paid to topological SLAM, which is bound to provide sufficiently accurate location estimates, while being significantly less computationally demanding. This paper intends to provide an introductory overview of the most prominent techniques that have been applied to topological SLAM in terms of feature detection, map matching, and map fusion.

Global optimisation algorithms for stereo dense depth map estimation have demonstrated how to outperform other stereo algorithms such as local methods or dynamic programming. The energy minimisation framework, using Markov random fields model and solved using graph cuts or belief propagation, has especially obtained good results. The main drawback of these methods is that, although they achieve accurate reconstruction, they are not suited for real-time applications. Subsampling the input images does not reduce the complexity of the problem because it also reduces the resolution of the output in the disparity space. Nonetheless, some real-time applications such as navigation would tolerate the reduction of the depth map resolutions (width and height) while maintaining the resolution in the disparity space (number of labels). In this study a new multiresolution energy minimisation framework for real-time robotics applications is proposed where a global optimisation algorithm is applied. A reduction by a factor R of the final depth map's resolution is considered and a speed of up to 50 times has been achieved. Using high-resolution stereo pair input images guarantees that a high resolution on the disparity dimension is preserved. The proposed framework has shown how to obtain real-time performance while keeping accurate results in the Middlebury test data set.

No abstract

No abstract

-During the last decade, the general-purpose computing on graphics processing units Graphics (GPGPU) has turned out to be a useful tool for speeding up many scientific calculations. Computer vision is known to be one of the fields with more penetration of these new techniques. This paper explores the advantages of using GPGPU implementation to speedup a genetic algorithm used for stereo refinement.The main contribution of this paper is analyzing which genetic operators take advantage of a parallel approach and the description of an efficient state-of-the-art implementation for each one. As a result, speed-ups close to x80 can be achieved, demonstrating to be the only way of achieving close to real-time performance.