Catálogo de publicaciones - libros

This paper presents a probabilistic approach for object localization which combines subspace clustering with the selection of discriminative clusters. Clustering is often a key step in object recognition and is penalized by the high dimensionality of the descriptors. Indeed, local descriptors, such as SIFT, which have shown excellent results in recognition, are high-dimensional and live in different low-dimensional subspaces. We therefore use a subspace clustering method called High-Dimensional Data Clustering (HDDC) which overcomes the curse of dimensionality. Furthermore, in many cases only a few of the clusters are useful to discriminate the object. We, thus, evaluate the discriminative capacity of clusters and use it to compute the probability that a local descriptor belongs to the object. Experimental results demonstrate the effectiveness of our probabilistic approach for object localization and show that subspace clustering gives better results compared to standard clustering methods. Furthermore, our approach outperforms existing results for the Pascal 2005 dataset.

Activity recognition consists of two fundamental tasks: tracking the features/objects of interest, and recognizing the activities. In this paper, we show that these two tasks can be integrated within the framework of a dynamical feedback system. In our proposed method, the recognized activity is continuously adapted based on the output of the tracking algorithm, which in turn is driven by the identity of the recognized activity. A non-linear, non-stationary stochastic dynamical model on the “shape” of the objects participating in the activities is used to represent their motion, and forms the basis of the tracking algorithm. The tracked observations are used to recognize the activities by comparing against a prior database. Measures designed to evaluate the performance of the tracking algorithm serve as a feedback signal. The method is able to automatically detect changes and switch between activities happening one after another, which is akin to segmenting a long sequence into homogeneous parts. The entire process of tracking, recognition, change detection and model switching happens recursively as new video frames become available. We demonstrate the effectiveness of the method on real-life video and analyze its performance based on such metrics as detection delay and false alarm.

Existing multi-kernel tracking methods are based on a forwards additive motion model formulation. However this approach suffers from the need to estimate an update matrix for each iteration. This paper presents a general framework that extends the existing approach and that allows to introduce a new inverse compositional formulation which shifts the computation of the update matrix to a one time initialisation step. The proposed approach thus reduces the computational complexity of each iteration, compared to the existing forwards approach. The approaches are compared both in terms of algorithmic complexity and quality of the estimation.

We present a generic framework to track shapes across large variations by learning non-linear shape manifold as overlapping, piecewise linear subspaces. We use landmark based shape analysis to train a Gaussian mixture model over the aligned shapes and learn a Point Distribution Model(PDM) for each of the mixture components. The target shape is searched by first maximizing the mixture probability density for the local feature intensity profiles along the normal followed by constraining the global shape using the most probable PDM cluster. The feature shapes are robustly tracked across multiple frames by dynamically switching between the PDMs. Our contribution is to apply ASM to the task of tracking shapes involving wide aspect changes and generic movements. This is achieved by incorporating shape priors that are learned over non-linear shape space and using them to learn the plausible shape space. We demonstrate the results on tracking facial features and provide several empirical results to validate our approach. Our framework runs close to real time at 25 frames per second and can be extended to predict pose angles using Mixture of Experts.

A successful approach for object tracking has been kernel based object tracking [1] by Comaniciu . The method provides an effective solution to the problems of representation and localization in tracking. The method involves representation of an object by a feature histogram with an isotropic kernel and performing a gradient based mean shift optimization for localizing the kernel. Though robust, this technique fails under cases of occlusion. We improve the kernel based object tracking by performing the localization using a based optimization. This makes the method resilient to occlusions. Another aspect related to the localization step is handling of scale changes by varying the bandwidth of the kernel. Here, we suggest a technique based on SIFT features [2] by Lowe to enable change of bandwidth of the kernel even in the presence of occlusion. We demonstrate the effectiveness of the techniques proposed through extensive experimentation on a number of challenging data sets.

Experiments in infant category formation indicate a strong role for temporal continuity and change in perceptual categorization. Computational approaches to model discovery in vision have traditionally focused on static images, with appearance features such as shape playing an important role. In this work, we consider integrating agent behaviors with shape for the purpose of agent discovery. Improved algorithms for video segmentation and tracking under occlusion enable us to construct models that characterize agents in terms of motion and interaction with other objects. We present a preliminary approach for discovering agents based on a combination of appearance and motion histories. Using uncalibrated camera images, we characterize objects discovered in the scene by their shape and motion attributes, and cluster these using agglomerative hierarchical clustering. Even with very simple feature sets, initial results suggest that the approach forms reasonable clusters for diverse categories such as people, and for very distinct clusters (animals), and performs above average on other classes.

This paper presents an approach of fusing the information provided by visible spectrum video with that of thermal infrared video to tackle video processing challenges such as object detection and tracking for increasing the performance and robustness of the surveillance system. An enhanced object detection strategy using gradient information along with background subtraction is implemented with efficient fusion based approach to handle typical problems in both the domains. An intelligent fusion approach using Fuzzy logic and Kalman filtering technique is proposed to track objects and obtain fused estimate according to the reliability of the sensors. Appropriate measurement parameters are identified to determine the measurement accuracy of each sensor. Experimental results are shown on some typical scenarios of detection and tracking of pedestrians.

Dynamic events comprise of spatiotemporal atomic units. In this paper we model them using a mixture model. Events are represented using a framework based on the Mixture of Factor Analyzers (MFA) model. It is to be noted that our framework is generic and is applicable for any mixture modelling scheme. The MFA, used to demonstrate the novelty of our approach, clusters events into spatially coherent mixtures in a low dimensional space. Based the observations that, (i) events comprise of varying degrees of spatial and temporal characteristics, and (ii) the number of mixtures determines the composition of these features, a method that incorporates models with varying number of mixtures is proposed. For a given event, the relative importance of each model component is estimated, thereby choosing the appropriate feature composition. The capabilities of the proposed framework are demonstrated with an application: recognition of events such as hand gestures, activities.

This paper presents an approach to identify the importance of different parts of a video sequence from the recognition point of view. It builds on the observations that: (1) events consist of more fundamental (or atomic) units, and (2) a discriminant-based approach is more appropriate for the recognition task, when compared to the standard modelling techniques, such as PCA, HMM, etc. We introduce which describe the usefulness of the fundamental units in distinguishing between events. We first extract actions to capture the fine characteristics of individual parts in the events. These actions are modelled and their usefulness in discriminating between events is estimated as a score. The score highlights the important parts (or actions) of the event from the recognition aspect. Applicability of the approach on different classes of events is demonstrated along with a statistical analysis.

Gesture segmentation is an extremely difficult task due to both the multitude of possible gesture variations in spatio-temporal space and the co-articulation of successive gestures. In this paper, a robust framework for this problem is proposed which has been used to segment out component gestures from a continuous stream of gestures using finite state machine and motion features in a vision based platform.