A dictionary learning approach to tracking

Finding efficient and powerful techniques for representing 2D and 3D entities is central in many areas, such as computer vision, graphics, CAD or data visualization. The large amount of data often involved and the small processing time required place strong demands on modeling techniques. This chapter gives an overview of the possibilities wavelets offer for object modeling, specially from the point of view of object recognition in computer vision environments. For the 2D case, techniques based on segmented contours or raw grey level information are described. For 3D data, both the cases of dense, regularly sampled volumetric information and sparse, irregularly sampled geometric surface information are considered. Ideas on how Bayesian methods may enhance wavelet based techniques for object representation and recognition are outlined.

This paper presents a spatial tracking system in a multiple camera environment. As a part of a general distributed sensing and system where the task of monitoring and tracking can be synchronized and divided between each of the modules, this multiple camera system covers a wider field of view, and detects moving objects by means of motion-based visual tracing. Based on camera calibration and geometry of multiple views, we propose some preliminary results for utilizing multiple cameras to reconstruct the spacial motion and trajectory of the object by establishing correlated mopping between the 2D projective images of the cameras.

In picture processing, picture segmentation, object classification and recognition is an impressive concept in present days because of increasing computer vision applications in real time applications. Based on required application, divide a picture into particular relevant clusters based on equivalent partitioning. Object identification is a minimum work to identify objects in a given picture or video with pixel values. The main representation of object classification from a given picture is to detect picture instances with respect to categorization of object. Object recognition is a complex task, when the pictures are poor quality, occlusion, and noise and in picture with background cutter sequences, this becomes more and efficient challenging task in picture processing. To appear object categorization and recognition effectively in pictures, there are several methods and approaches were introduced to solve object representation effectively. In this paper we review the different techniques of contextual feature extraction with respect to picture information of object classification, categorization and recognition of scalability and optimizations in real time picture processing applications. Our research gives better review of different approaches in picture categorization.

A survey of moment-based techniques for unoccluded object representation and recognition

Due to the large intraclass variances and complicated object distribution, recognizing objects with complex appearances and arbitrary orientations has been an active research topic and a challenging task in remote sensing fields. In this article, we formulate object recognition as a high-level feature-learning problem, and a novel supervised method is proposed to learn high-level feature representations from high-resolution remote sensing images for object recognition. Our method simultaneously and coherently achieves high-level feature learning and classifier training, which improves the recognition performance. Two constraints that enforce the label consistencies of group images and label consistencies of single images are introduced in a deep learning framework to obtain the high-level feature space. The high-level feature and a multiclass linear classifier are finally learned by an effective optimization algorithm. Experimental results demonstrate the superior performance of the proposed method over many state-of-the-art techniques in object recognition.

Most fMRI studies of the brain’s intrinsic functional connectivity (FC) have assumed that this is static; however, it is now clear that it changes over time. This is particularly relevant in epilepsy, which is characterized by a continuous interchange between epileptic and normal brain states associated with the occurrence of epileptic activity. Interestingly, recurrent states of dynamic FC (dFC) have been found in fMRI data using unsupervised learning techniques, assuming either their sparse or non-sparse combination. Here, we propose an l1-norm regularized dictionary learning (l1-DL) approach for dFC state estimation, which allows an intermediate and flexible degree of sparsity in time, and demonstrate its application in the identification of epilepsy-related dFC states using simultaneous EEG-fMRI data. With this l1-DL approach, we aim to accommodate a potentially varying degree of sparsity upon the interchange between epileptic and non-epileptic dFC states. The simultaneous recording of the EEG is used to extract time courses representative of epileptic activity, which are incorporated into the fMRI dFC state analysis to inform the selection of epilepsy-related dFC states. We found that the proposed l1-DL method performed best at identifying epilepsy-related dFC states, when compared with two alternative methods of extreme sparsity (k-means clustering, maximum; and principal component analysis, minimum), as well as an l0-norm regularization framework (l0-DL), with a fixed amount of temporal sparsity. We further showed that epilepsy-related dFC states provide novel insights into the dynamics of epileptic networks, which go beyond the information provided by more conventional EEG-correlated fMRI analysis, and which were concordant with the clinical profile of each patient. In addition to its application in epilepsy, our study provides a new dFC state identification method of potential relevance for studying brain functional connectivity dynamics in general.

Dictionary learning (DL) for sparse coding has been widely applied in the field of computer vision. Many DL approaches have been developed recently to solve pattern classification problems and have achieved promising performance. In this paper, to improve the discriminability of the popular dictionary pair learning (DPL) algorithm, we propose a new method called discriminative dictionary pair learning (DDPL) for image classification. To achieve the goal of signal representation and discrimination, we impose the incoherence constraints on the synthesis dictionary and the low-rank regularization on the analysis dictionary. The DDPL method ensures that the learned dictionary has the powerful discriminative ability and the signals are more separable after coding. We evaluate the proposed method on benchmark image databases in comparison with existing DL methods. The experimental results demonstrate that our method outperforms many recently proposed dictionary learning approaches.

A novel dictionary learning approach for multi-modality medical image fusion

A novel technique for object recognition and localization within a 3D point cloud has been developed, by constructing a likelihood function for the pose vector of a known model in a measured scene. The function is based on surface features in the model and corresponding surface features detected in the scene. Using an optimization algorithm, the maximum of the function was found, corresponding to the 6 degree of freedom (DOF) pose of the model within the scene even in the presence of significant clutter.

Object Recognition has been a field in Computer Vision research, which is far from being solved when it comes to localizing the object of interest in an unconstrained environment, captured from different viewing angles. Lack of benchmark datasets clogs the progress in this field since the last decade, barring the subset of a single dataset, alias the Office dataset, which attempted to boost research in the field of pose-invariant detection and recognition of portable object in unconstrained environment. A new challenging object dataset with 30 categories has been proposed with a vision to boost the performances of the task of object recognition for portable objects, thus enhancing the study of cross domain adaptation, in conjunction to the Office dataset. Images of various hand-held objects are captured by the primary camera of a smartphone, where they are photographed under unconstrained environment with varied illumination conditions at different viewing angles. The monte-carlo object detection and recognition has been performed for the proposed dataset, facilitated by existing state-of-the-art transfer learning techniques for cross-domain recognition of objects. The baseline accuracies for existing Domain Adaptation methods, published recently, are also presented in this paper, for the kind perusal of the researchers. A new technique has also been proposed based on the activation maps of the AlexNet to detect objects, alongwith a Generative Adversarial Network (GAN) based Domain Adaptation technique for Object Recognition.

A novel method of three-dimensional (3-D) object recognition is proposed. Several projections of a 3-D target are recorded under white-light illumination and fused into a single complex two-dimensional function. After proper filtering, the resulting function is coded into a computer-generated hologram. When this hologram is coherently illuminated, a correlation space is reconstructed such that light peaks indicate the existence and locations of true targets in the observed 3-D scene. Experimental results and comparisons with results of another 3-D object recognition technique are presented.

Many human activities are based on visual recognition of a scene. Development of scene recognition capability on computer is essential to the development of an intelligent robot.This paper proposes a three‐dimensional (3‐D) object recognition system using functional models. In conventional geometric model‐based recognition, one reference model must be provided for each possible object shape in the same category. This paper proposes a functional model that will allow flexibility in representation of objects belonging to the same category.Experiments were conducted to provide measurement data for demonstrating this functional model‐based object recognition technique. Scene segmentation uses both gray scale image and range image to provide superior segmentation results. The performance of a functional model‐based method for recognizing objects with different shapes that belong to the same functional category is demonstrated.

Classical dictionary learning algorithms that rely on a single source of information have been successfully used for the discriminative tasks. However, exploiting multiple sources has demonstrated its effectiveness in solving challenging real-world situations. We propose a new framework for feature fusion to achieve better classification performance as compared to the case where individual sources are utilized. In the context of multimodal data analysis, the modality configuration induces a strong group/coupling structure. The proposed method models the coupling between different modalities in space of sparse codes while at the same time within each modality a discriminative dictionary is learned in an all-vs-all scheme whose class-specific sub-parts are non-correlated. The proposed dictionary learning scheme is referred to as the multimodal weighted dictionary learning (MWDL). We demonstrate that MWDL outperforms state-of-the-art dictionary learning approaches in various experiments.

We present in this paper a real-time method for visual categorization to do robot grasping. We describe an object database with SURF feature points which we quantify with the Kmeans clustering algorithm to make visual words. Then, we train a Support Vector Machine classifier having as entries the distribution of the bag of features extracted earlier. Likewise, we do object recognition using the SVM algorithm. The real-time implementation is done with the OpenCV GPU. The application aims to pick up an object and drop it using our robot manipulator equipped with a camera using our visual system. Finally, the results of our experiments of the object recognition show an average of recognition between 95% and 100%.

A robot vision system can be called humanoid if it possesses an oculomotor system similar to human eyes and if it is capable to simultaneously acquire and process images of varying resolution. Designers of a number of humanoid robots attempted to mimic the foveated structure of the human eye. Foveation is useful because, firstly, it enables the robot to monitor and explore its surroundings in images of low resolution, thereby increasing the efficiency of the search process, and secondly, it makes it possible to simultaneously extract additional information – once the area of interest is determined – from higher resolution foveal images that contain more detail. There are several visual tasks that can benefit from foveated vision. One of the most prominent among them is object recognition. General object recognition on a humanoid robot is difficult because it requires the robot to detect objects in dynamic environments and to control the eye gaze to get the objects into the fovea and to keep them there. Once these tasks are accomplished, the robot can determine the identity of the object by processing foveal views. Approaches proposed to mimic the foveated structure of biological vision systems include the use of two cameras per eye (Atkeson et al., 2000; Breazeal et al., 2001; Kozima & Yano, 2001; Scassellati, 1998) (Cog, DB, Infanoid, Kismet, respectively), i. e. a narrow-angle foveal camera and a wide-angle camera for peripheral vision; lenses with space-variant resolution (Rougeaux & Kuniyoshi, 1998) (humanoid head ESCHeR), i. e. a very high definition area in the fovea and a coarse resolution in the periphery; and space-variant log-polar sensors with retina-like distribution of photo-receptors (Sandini &Metta, 2003) (Babybot). It is also possible to implement log-polar sensors by transforming standard images into log-polar ones (Engel et al., 1994), but this approach requires the use of high definition cameras to get the benefit of varying resolution. Systems with zoom lenses have some of the advantages of foveated vision, but cannot simultaneously acquire wide angle and high resolution images. Our work follows the first approach (see Fig. 1) and explores the advantage of foveated vision for object recognition over standard approaches, which use equal resolution across the visual field. While log-polar sensors are a closer match to biology, we note that using two cameras per eye can be advantageous because cameras with standard chips can be utilized. This makes it possible to equip a humanoid robot with miniature cameras (lipstick size and 1