Stjepan Rajko

Visibility-Based Pursuit-Evasion in an Unknown Planar Environment

We address an on-line version of the visibility-based pursuit-evasion problem. We take a minimalist approach in modeling the capabilities of a pursuer robot. A point pursuer moves in an unknown, simplyconnected, piecewise-smooth planar environment, and is given the task of locating any unpredictable, moving evaders that have unbounded speed. The evaders are assumed to be points that move continuously. To solve the problem, the pursuer must for each target have an unobstructed view of it at some time during execution. The pursuer is equipped with a range sensor that measures the direction of depth discontinuities, but cannot provide precise depth measurements. All pursuer control is specified either in terms of this sensor or wall-following movements. The pursuer does not have localization capability or perfect control. We present a complete algorithm that enables the limited pursuer to clear the same environments that a pursuer with a complete map, perfect localization, and perfect control can clear (under certain general position assumptions). Theoretical guarantees that the evaders will be found are provided. The resulting algorithm to compute this strategy has been implemented in simulation. Results are shown for several examples. The approach is efficient and simple enough to be useful towards the development of real robot systems that perform visual searching.

Real-time Gesture Recognition with Minimal Training Requirements and On-line Learning

In this paper, we introduce the semantic network model (SNM), a generalization of the hidden Markov model (HMM) that uses factorization of state transition probabilities to reduce training requirements, increase the efficiency of gesture recognition and on-line learning, and allow more precision in gesture modeling. We demonstrate the advantages both formally and experimentally, using examples such as full-body multimodal gesture recognition via optical motion capture and a pressure sensitive floor, as well as mouse/pen gesture recognition. Our results show that our algorithm performs much better than the traditional approach in situations where training samples are limited and/or the precision of the gesture model is high.

A pursuit-evasion BUG algorithm

We consider the problem of searching for an unpredictable moving target, using a robot that lacks a map of the environment, lacks the ability to construct a map, and has imperfect navigation ability. We present a complete algorithm, which yields a motion strategy for the robot that guarantees the elusive target will be detected, if such a strategy exists. It is assumed that the robot has an omnidirectional sensing device that is used to detect moving targets and also discontinuities in depth data in a 2D environment. We also show that the robot has the same problem solving power as a robot that has a complete map and perfect navigation abilities. The algorithm has been implemented in simulation, and some examples are shown.

The design of a pressure sensing floor for movement-based human computer interaction

This paper addresses the design of a large area, high resolution, networked pressure sensing floor with primary application in movement-based human-computer interaction (M-HCI). To meet the sensing needs of an M-HCI system, several design challenges need to be overcome. Firstly, high frame rate and low latency are required to ensure real-time human computer interaction, even in the presence of large sensing area (for unconstrained movement in the capture space) and high resolution (to support detailed analysis of pressure patterns). The optimization of floor system frame rate and latency is a challenge. Secondly, in many cases of M-HCI there are only a small number of subjects on the floor and a large portion of the floor is not active. Proper data compression for efficient data transmission is also a challenge. Thirdly, locations of disjoint active floor regions are useful features in many M-HCI applications. Reliable clustering and tracking of active disjoint floor regions poses as a challenge. Finally, to allow M-HCI using multiple communication channels, such as gesture, pose and pressure distributions, the pressure sensing floor needs to be integrable with other sensing modalities to create a smart multimodal environment. Fast and accurate alignment of floor sensing data in space and time with other sensing modalities is another challenge. In our research, we fully addressed the above challenges. The pressure sensing floor we developed has a sensing area of about 180 square feet, with a sensor resolution of 6.25 sensels/in2. The system frame rate is up to 43 Hz with average latency of 25 ms. A simple but efficient data compression scheme is in place. We have also developed a robust clustering and tracking procedure for disjoint active floor regions using the mean-shift algorithm. The pressure sensing floor can be seamlessly integrated with a marker based motion capture system with accurate temporal and spatial alignment. Furthermore, the modular and scalable structure of the sensor floor allows for easy installation to real rooms of irregular shape. The pressure sensing floor system described in this paper forms an important stepping stone towards the creation of a smart environment with context aware data processing algorithms which finds extensive applications beyond M-HCI, e.g. diagnosing gait pathologies and evaluation of treatment.

View-invariant full-body gesture recognition via multilinear analysis of voxel data

This paper presents a gesture recognition framework using voxel data obtained through visual hull reconstruction from multiple cameras. View-invariant pose descriptors are extracted by projecting voxel data onto a low dimensional pose coefficient space using multilinear analysis. Gestures are then treated as sequences of pose descriptors and represented by hidden Markov models for gesture recognition. Promising results have been obtained using a public data set containing 11 single-person gestures and another data set including seven two-people cooperative dance gestures.

A dynamic Bayesian approach to computational Laban shape quality analysis

Laban movement analysis (LMA) is a systematic framework for describing all forms of human movement and has been widely applied across animation, biomedicine, dance, and kinesiology. LMA (especially Effort/Shape) emphasizes how internal feelings and intentions govern the patterning of movement throughout the whole body. As we argue, a complex understanding of intention via LMA is necessary for human-computer interaction to become embodied in ways that resemble interaction in the physical world. We thus introduce a novel, flexible Bayesian fusion approach for identifying LMA Shape qualities from raw motion capture data in real time. The method uses a dynamic Bayesian network (DBN) to fuse movement features across the body and across time and as we discuss can be readily adapted for low-cost video. It has delivered excellent performance in preliminary studies comprising improvisatory movements. Our approach has been incorporated in Response, a mixed-reality environment where users interact via natural, full-body human movement and enhance their bodily-kinesthetic awareness through immersive sound and light feedback, with applications to kinesiology training, Parkinsons patient rehabilitation, interactive dance, and many other areas.

HMM parameter reduction for practical gesture recognition

We examine in detail some properties of gesture recognition models which utilize a reduced number of parameters and lower algorithmic complexity compared to traditional hidden Markov models. We show that the reduced parameter models are comparable to standard HMM-based gesture recognition models in their ability to effectively model gestures, and in some cases superior when training data is limited. We also show that in order to effectively differentiate similar gestures, a gesture recognition model must utilize a large number of states, a scenario which can only be adequately handled by reducer parameter methods to maintain real-time speeds.

Space and time optimal parallel sequence alignments

We present the first space and time optimal parallel algorithm for the pairwise sequence alignment problem, a fundamental problem in computational biology. This problem can be solved sequentially in O(mn) time and O(m+n) space, where m and n are the lengths of the sequences to be aligned. The fastest known parallel space-optimal algorithm for pairwise sequence alignment takes optimal O(m+n/p) space but suboptimal O((m+n)2/p) time, where p is the number of processors. On the other hand, the most space economical time-optimal parallel algorithm takes O(mn/p) time but O(m+n/p) space. We close this gap by presenting an algorithm that achieves both time and space optimality, i.e. requires only O(m+n/p) space and O(mn/p) time. We also present an experimental evaluation of the proposed algorithm on an IBMxSeries cluster

A hybrid HMM/DPA adaptive gesture recognition method

We present a hybrid classification method applicable to gesture recognition. The method combines elements of Hidden Markov Models (HMM) and various Dynamic Programming Alignment (DPA) methods, such as edit distance, sequence alignment, and dynamic time warping. As opposed to existing approaches which treat HMM and DPA as either competing or complementing methods, we provide a common framework which allows us to combine ideas from both HMM and DPA research. The combined approach takes on the robustness and effectiveness of HMMs and the simplicity of DPA approaches. We have implemented and successfully tested the proposed algorithm on various gesture data.

Design Optimization of Pressure Sensing Floor for Multimodal Human-Computer Interaction

Human-computer interaction is a discipline concerned with the design, evaluation and implementation of interactive computing systems for human use. Humans communicate with each other, intentionally or unintentionally, using various interpersonal communication modes such as static and dynamic full-body, limb, and hand gestures, facial expressions, speech and sounds, and haptics, just to name a few. It is natural to design human-computer interaction systems with which users can communicate using these interpersonal communication modes. To this end, multimodal human-computer interaction (MMHCI) systems are receiving increasing attention recently. An overview of the recent advances of MMHCI can be found in (Jaimes and Sebe, 2007). Our research mainly focuses on movement analysis based on visual and pressure sensing for movement based MMHCI, which read the movement of user(s), and respond accordingly through real-time visual and audio feedback. Such movement based MMHCI systems have immediate applications in a number of areas with significant impact on our daily lives, including biomedical, e.g. rehabilitation of stroke patients (Chen, et al., 2006), culture and arts, e.g. studying patterns and cues in complex dance performances, and interactive dance performances (Qian, et al., 2004), K-12 education, e.g. collaborative and embodied learning (Birchfield, et al., 2006), sports (e.g. analyzing and improving athletic performance based on weight distributions), and security (e.g. movement based smart surveillance systems), just to name a few.