This paper presents an accurate and robust processing pipeline for merging tridimensional datasets related to railway contexts and thus producing a comprehensive 3D model of the monitored scenario. The method is made of multiple modules able to detect the rail tracks and achieve consecutive point cloud registration. A preliminary stage is aimed at filtering out those outlier points and selecting only specific regions of interest from the point cloud. Afterwards, the procedure detects the 3D points belonging to the rail tracks, which can be considered as good candidates for attaining the final point cloud registration. A local analysis for each 3D point is performed by considering a parallelepiped-shaped voxel opportunely centered at the point under investigation. The evaluation of the spatial distributions of points inside the considered volume voxel is performed in order to establish if a seed point lies on the rail head. Further checks enable to reject false candidate points from previous steps by taking advantage of the knowledge about the rail track gauge. Finally, a hierarchical clustering completes the extraction of potential rails. The registration module uses the IterativeClosest-Point method, combined with an algorithm that iteratively reduces the overlapping regions between two consecutive point clouds, for merging the data by using the rail points. The methodology is validated on two different datasets collected by using a stereo camera developed at our laboratory. Final outcomes prove as the proposed approach enables to attain robust and accurate global 3D registration in railway contexts.
We propose a method for solving one of the significant open issues in computer vision: material recognition. A time-of-flight range camera has been employed to analyze the characteristics of different materials. Starting from the information returned by the depth sensor, different features of interest have been extracted using transforms such as Fourier, discrete cosine, Hilbert, chirp-z, and Karhunen–Loève. Such features have been used to build a training and a validation set useful to feed a classifier (J48) able to accomplish the material recognition step. The effectiveness of the proposed methodology has been experimentally tested. Good predictive accuracies of materials have been obtained. Moreover, experiments have shown that the combination of multiple transforms increases the robustness and reliability of the computed features, although the shutter value can heavily affect the prediction rates.
This paper introduces a hybrid methodology that ensemble genetic algorithms and Support Vector Machine (SVM) in order to evolve optimal subsets of Gabor filters for efficient pattern classification. ALthough some filter design procedure are available for Gabor filters, high computations are needed and the efficiency of design is dependent on the particualr Gabor filter subset. In this paper to reduce the computational cost and improve the performance, a GA is used to search the space of all possible subsets of a large pool of Gabor candidate filters. The classification performance of SVM, an unknown data, together with filtering cost are used as measure of fitness that is used as feedback by GA to evolve better Gabor filter sets. This assembled system iterates until filters subset is found with a satisfactory classification performance and a significant reduced filters number.
In the last years the detection and classification of surface defects of material is assuming great importance. Visual inspection can help to increase the product quality and, in particular context, the maintenance of products. The railway infrastructure is a particular field in which the periodical surface inspection of rolling plane can help an operator to prevent critical situation. We use a Gabor filter to emphasize the image regions with grey level variation. The Gabor filter h(x,y) is characterized by a frequency F, direction (theta) and parameter (sigma) . We have selected experimentally four filters with directions 0, (pi) /4, (pi) /2 and (pi) 3/4 with F equals (root)2/8 cycle/pixel and (sigma) equals 2. The problem of detection and classification is a crucial part of our work because cannot be defined an exhaustive training set of defect and no-defect images. It is necessary a method able to self-learn changes. Investigating about this problem we propose in the paper a novel Self Organized Map (SOM) network, appropriately modified, for detection and classification of rail defects. The proposed SOM network learns to classify input vectors according to how they are grouped in the input space. So, SOM learns both the distribution and topology of the input vectors belonging to the training set. During the training phase, the neurons in the layer of an SOM form some cluster or bubble representing the input training with minimum distance among them. The novelty is to modify the SOM network in order to learn continuously during the test phase.
This paper describes a technique for detection of unknown obstacles using a stereo pair of TV cameras, for mobile robot navigational purpose. Three-dimensional information is recovered by matching segments. Moreover a feature grouping technique is used to produce a coarse obstacle reconstruction, but enough for detecting the free space (without obstacles) in the environment. The advantage is to use a such reconstructed obstacle map is twofold: higher resolution than map obtained by active sensors such as ultrasonics and, moreover, the obstacles are detected from far than active sensors. Results on experimental stereo images, acquired in our laboratory, are presented in order to illustrate the reliability of the technique.
Accurate position estimation is a fundamental requirement for mobile robot navigation. The positioning problem consists of keeping in real-time a reliable estimate of the robot location with respect to a reference frame in the environment. A fast landmark-based position estimation method is presented in this paper. The technique combines orientation of the mobile robot from a heading sensor (a compass) with observations of landmarks from a vision sensor (a CCD camera). Knowing the position of the landmarks in a fixed coordinate system and the orientation of the optical axis of the camera it's possible to recover the robot position by simple geometric considerations. The experiments made in our laboratory demonstrate the reliability of the method and suggest its applicability in the context of autonomous robot navigation.
Self-location is the capability of a mobile robot to determine its position in the environment referring to absolute landmarks. The possibility to use natural visual landmarks for self-location augments the autonomy and the flexibility of mobile vehicles. In this paper the use of junctions, detected in real images, as landmarks is proposed. The use of visual cues means that problems regarding variations of perspective and scale must be resolved. We propose to formulate the junction recognition as a graph matching problem and resolved using standard methods. Experimental results are shown on real contexts.
Low-level navigation for autonomous vehicles can be accomplished efficiently by a behavioral-based approach that involves the simultaneous execution of independent sub-tasks seen as primitive behaviors. Each behavior maps sensory data into control commands in a reactive way, with no need of internal representations. A useful tool for realizing such a direct mapping is fuzzy logic, that allows the production of control rules by either manual programming or automatic learning. In prospect of implementing an articulated control system including all the low-level behaviors of navigation, this paper focuses on the problem of obtaining an efficient and robust fuzzy controller performing a single behavior and presents a method for minimizing the number of rules of a fuzzy controller developed for driving a TRC Labmate based vehicle along the wall on its right-hand side. Fuzzy rules, that map ultrasonic sensor readings onto steering velocity values, are learned automatically from training data collected during operator-driven runs of the vehicle. In addition, we address the problem of defining an appropriate performance function, that may be useful for evaluating the influence of the rule base reduction on the overall behavior of the vehicle during navigation, but also for estimating the quality of a control rule, in order to adapt rules on- line. Results of an experimental comparison between the original fuzzy wall-follower and its optimized version are reported.
Our goal is to match primitives of a pair of images, thereby solving the correspondence problem, in order to estimate depths of 3D scene points from the relative distance between matched features. We propose a feature-based approach to solve the correspondence problem by minimizing an appropriate energy function where constraints on radiometric similarity and projective geometric invariance of coplanar points are defined. The method can be seen as a correlation based approach which takes into account the projective invariance of coplanar points in computing the optimal matches.
Navigation in dynamic indoor environments requires a mobile vehicle to follow the planned path while avoiding unexpected obstacles eventually met along it. In this paper an attempt of designing a path planar using a computational model suitable for fast implementation on special purpose hardware is presented, in which the automatic modeling of the scene and its continuous updating are accomplished by means of a recursive ultrasonic-based obstacle avoidance system. From this model a graph representing all the possible paths for the robot in the free-space is built using well known methodologies (configuration space, generalized cones). The task of searching for the shortest path in this graph is solved by means of a neural network based on the Hopfield model, that represents an interesting alternative to classical techniques as A*. A major advantage of this neural approach is the parallel nature of the resulting network that allows a rapid convergence to a solution when implemented in hardware. Simulation results are shown to illustrate the performance of the Hopfield path planner.
In this work we consider the application context of planar passive navigation in which the visual control of locomotion requires only the direction of translation and not the full set of motion parameters. If the temporally changing optic array is represented as a vector field of optical velocities, the vectors form a radial pattern emanating from a center point, called the focus of expansion (FOE), representing the heading direction. The FOE position is independent of the distances of world surfaces and doesn't require assumptions about surface shape and smoothness. We investigate the performance of an artificial neural network for the computation of the image position of the FOE of an optical flow field induced by an observer translation relative to a static environment. The network is characterized by a feed forward architecture and is trained by a standard supervised back-propagation algorithm which receives as input the pattern of points where the lines generated by 2D vectors are projected using the Hough transform. We present results obtained on test set of synthetic noisy optical flows and on optical flows computed from real image sequences.
In this paper a method for the estimation of the heading direction and of the time-of-collision of a moving vehicle is presented. The assumption that the motion can be described as a prevalence of translation motion is used to reduce the optic flow equations to a linear version. In this case 2D motion field assumes a radial shape with vectors directions intersecting in a point called focus of expansion. In the presented method a sparse linear optic flow map is derived in the most relevant and reliable areas of the image. These estimations are then used to derive information about 3D motion of the vehicle. Results on synthetic and real time-varying sequence are presented.
Many of visual navigation strategies for an autonomous mobile robot are landmark based. A vehicle to determine its position needs to refer to absolute references in the environment, so landmarks are required to be invariant for rotation, translation, scale and perspective. A straightforward alternative is to be able to characterize invariantly the context where landmarks are placed. In this paper, we show as a neural network appropriately trained, is able to recognize context where landmarks are located in the scene. The early results seem to be interesting.
The work here presented proposes an iterative refining algorithm to build a sequence of convex functionals based essentially on a weak thin-plate under tension model for smoothing. This algorithm is applied to structure estimation and edge detection problems. The sequence of functionals is obtained by modifying continuously and iteratively continuous non-binary line processes which control regularity of the surface. This is done by comparing the smoothed estimation with initial data (sparse or dense), that is evaluating signal-to-noise characteristics.
A simple vision based system to perform a contactless coupling, rather than a hardware hook is suggested. The method facilitates automatic convoy formation and management. The vision technique uses 3 landmarks and a CCD camera. Using a geometric approach that capitalizes the excellent angular resolution of CCD cameras, the position and orientation of the camera are estimated in respect to the landmarks. Experimental results show the reliability of the technique in operating contexts.
This paper deals with a data fusion technique for depth reconstruction which integrates regularization by variational methods with stochastic optimization based on Kalman filtering. A framework for the fusion of multiple regularized depth maps is proposed for on-line integration of many views of the visible scene. This kind of approach has some advantages in respect with similar ones, as it is stressed widely in the paper. It does not use optical flow, camera modeling or an explicit motion equation and can be used to fuse stochastically both sparse or dense depth data, obtaining reliable estimates in the whole image domain.
Detecting unexpected obstacles and avoiding collisions is an important task for any autonomous mobile robot. In this paper we describe an approach using a sonar-based system that we have used in our indoor autonomous mobile system. The logical design of this system is shown, followed by a description of how it builds a knowledge of the environment. The information collected of the environment can be used for many applications like real-time obstacle avoidance, environment learning, position estimation. This method builds up two kind of maps: a occupancy grid which contains the probability value of each cell to be occupied and a orientation map which contains the expected orientation of the surface of each cell in the occupancy grid. Methods for filtering raw sensor data before using it for map generation together with experimental results are shown.
Recently there has been an increasing interest in the development of a robot system capable of moving inside buildings. The problem of continuously establishing the position of a vehicle is fundamental in goal-oriented navigation. In this work we propose a vision system for an autonomous vehicle capable of determining its real position with respect to the planned one. A set of control points is given in a fixed coordinate system. If these points are identified in the image plane, the location from which the image has been obtained, and therefore the vehicle position, can be determined.
Depth and orientation information are important cues for the reconstruction of three-dimensional surfaces in computer vision. The statistical fusion of data obtained by slightly different views of the same scene is studied as a way for improving the accuracy and reliability of the data and consequent result of the integration step.
In this work we propose a segmentation algorithm working on surface normal information. An edge-based coarse segmentation map is computed by the detection of discontinuities in surface orientations. A region-based segmentation is generated by the analysis of surface curvature. A decision making process produces the final segmentation.