Mario Sznaier

Person Re-Identification Using Kernel-Based Metric Learning Methods

Re-identification of individuals across camera networks with limited or no overlapping fields of view remains challenging in spite of significant research efforts. In this paper, we propose the use, and extensively evaluate the performance, of four alternatives for re-ID classification: regularized Pairwise Constrained Component Analysis, kernel Local Fisher Discriminant Analysis, Marginal Fisher Analysis and a ranking ensemble voting scheme, used in conjunction with different sizes of sets of histogram-based features and linear, χ 2 and RBF-χ 2 kernels. Comparisons against the state-of-art show significant improvements in performance measured both in terms of Cumulative Match Characteristic curves (CMC) and Proportion of Uncertainty Removed (PUR) scores on the challenging VIPeR, iLIDS, CAVIAR and 3DPeS datasets.

Suboptimal control of linear systems with state and control inequality constraints

A suboptimal controller based upon on-line quadratic programming is described. Theoretical results are presented to show that such a controller is optimal under the assumption that there are no constraints on the computation time. Finally, an implementation of a suboptimal controller that takes such constraints into account is described.

The Way They Move: Tracking Multiple Targets with Similar Appearance

We introduce a computationally efficient algorithm for multi-object tracking by detection that addresses four main challenges: appearance similarity among targets, missing data due to targets being out of the field of view or occluded behind other objects, crossing trajectories, and camera motion. The proposed method uses motion dynamics as a cue to distinguish targets with similar appearance, minimize target mis-identification and recover missing data. Computational efficiency is achieved by using a Generalized Linear Assignment (GLA) coupled with efficient procedures to recover missing data and estimate the complexity of the underlying dynamics. The proposed approach works with track lets of arbitrary length and does not assume a dynamical model a priori, yet it captures the overall motion dynamics of the targets. Experiments using challenging videos show that this framework can handle complex target motions, non-stationary cameras and long occlusions, on scenarios where appearance cues are not available or poor.

Heuristically enhanced feedback control of constrained discrete-time linear systems

Recent advances in computer technology have spurred new interest in the use of feedback controllers based upon on-line minimization for the control of constrained linear systems. Still the use of computers in the feedback loop has been hampered by the fact that the amount of time available for computation in most sampled data systems is not enough to achieve a complete solution using conventional algorithms. Several “ad hoc” techniques have been proposed, but their applicability is restricted by the lack of supporting theory. In this paper we present a theoretical framework to analyze the stability of the closed-loop system resulting from the use of on-line optimization in the feedback loop. Using these results we show that a suboptimal algorithm, based upon the use of heuristic search techniques, yields asymptotically stable systems, provided that enough computation power is available to solve at each sampling interval an optimization problem considerably simpler than the original. The controller presented in this paper is valuable for situations where the customary approaches of using Pontryagins minimum principle or storing a family of extremal curves are not applicable due to limitations in the computational resources available.

Persistent disturbance rejection via static-state feedback

In contrast with /spl Hscr//sub /spl infin// and /spl Hscr//sub 2/ control theories, the problem of persistent disturbance rejection (l/sup 1/ optimal control) leads to dynamic controllers, even when the states of the plant are available for feedback. Using viability theory, Shamma showed (1993), in a nonconstructive way, that in the state-feedback case the same performance achieved by any dynamic linear time-invariant controller can be achieved using memoryless nonlinear state feedback. In this paper we give an alternative, constructive proof of these results for discrete- and continuous-time systems. The main result of the paper shows that in both cases, the l/sup 1/ norm achieved by any stabilizing state-feedback linear dynamic controller can be also achieved using a memoryless variable structure controller. >

Cross-view activity recognition using Hankelets

Human activity recognition is central to many practical applications, ranging from visual surveillance to gaming interfacing. Most approaches addressing this problem are based on localized spatio-temporal features that can vary significantly when the viewpoint changes. As a result, their performances rapidly deteriorate as the difference between the viewpoints of the training and testing data increases. In this paper, we introduce a new type of feature, the “Hankelet” that captures dynamic properties of short tracklets. While Hankelets do not carry any spatial information, they bring invariant properties to changes in viewpoint that allow for robust cross-view activity recognition, i.e. when actions are recognized using a classifier trained on data from a different viewpoint. Our experiments on the IXMAS dataset show that using Hanklets improves the state of the art performance by over 20%.

Dynamic Appearance Modeling for Human Tracking

Dynamic appearance is one of the most important cues for tracking and identifying moving people. However, direct modeling spatio-temporal variations of such appearance is often a difficult problem due to their high dimensionality and nonlinearities. In this paper we present a human tracking system that uses a dynamic appearance and motion modeling framework based on the use of robust system dynamics identification and nonlinear dimensionality reduction techniques. The proposed system learns dynamic appearance and motion models from a small set of initial frames and does not require prior knowledge such as gender or type of activity. The advantages of the proposed tracking system are illustrated with several examples where the learned dynamics accurately predict the location and appearance of the targets in future frames, preventing tracking failures due to model drifting, target occlusion and scene clutter.

H2/H∞ filtering theory and an aerospace application

In many filtering problems of practical interest, some of the noise signals satisfy the assumptions of H 2 (KalmanBucy) filtering, while others can be more accurately modeled as bounded energy signals (hence more amenable to an H, filtering approach). These problems may be addressed by considering a mixed H 2 I ? i , filtering problem. In this paper we present a novel theory which solves the mixed problem ezac t ly and in a computationally efficient way. The applicability of the theory is illustrated by designing a filter to estimate the states of an aircraft flying through a downburst.

A Sparsification Approach to Set Membership Identification of Switched Affine Systems

This paper addresses the problem of robust identification of a class of discrete-time affine hybrid systems, switched affine models, in a set membership framework. Given a finite collection of noisy input/output data and some minimal a priori information about the set of admissible plants, the objective is to identify a suitable set of affine models along with a switching sequence that can explain the available experimental information, while minimizing either the number of switches or subsystems. For the case where it is desired to minimize the number of switches, the key idea of the paper is to reduce this problem to a sparsification form, where the goal is to maximize sparsity of a suitably constructed vector sequence. Our main result shows that in the case of ℓ∞ bounded noise, this sparsification problem can be exactly solved via convex optimization. In the general case where the noise is only known to belong to a convex set N, the problem is generically NP-hard. However, as we show in the paper, efficient convex relaxations can be obtained by exploiting recent results on sparse signal recovery. Similarly, we present both a sparsification formulation and a convex relaxation for the (known to be NP hard) case where it is desired to minimize the number of subsystems. These results are illustrated using two non-trivial problems arising in computer vision applications: video-shot and dynamic texture segmentation.

Activity recognition using dynamic subspace angles

Cameras are ubiquitous everywhere and hold the promise of significantly changing the way we live and interact with our environment. Human activity recognition is central to understanding dynamic scenes for applications ranging from security surveillance, to assisted living for the elderly, to video gaming without controllers. Most current approaches to solve this problem are based in the use of local temporal-spatial features that limit their ability to recognize long and complex actions. In this paper, we propose a new approach to exploit the temporal information encoded in the data. The main idea is to model activities as the output of unknown dynamic systems evolving from unknown initial conditions. Under this framework, we show that activity videos can be compared by computing the principal angles between subspaces representing activity types which are found by a simple SVD of the experimental data. The proposed approach outperforms state-of-the-art methods classifying activities in the KTH dataset as well as in much more complex scenarios involving interacting actors.