Raúl Mohedano

Robust 3D people tracking and positioning system in a semi-overlapped multi-camera environment

People positioning and tracking in 3D indoor environments are challenging tasks due to background clutter and occlusions. Current works are focused on solving people occlusions in low-cluttered backgrounds, but fail in high-cluttered scenarios, specially when foreground objects occlude people. In this paper, a novel 3D people positioning and tracking system is presented, which shows itself robust to both possible occlusion sources: static scene objects and other people. The system holds on a set of multiple cameras with partially overlapped fields of view. Moving regions are segmented independently in each camera stream by means of a new background modeling strategy based on Gabor filters. People detection is carried out on these segmentations through a template-based correlation strategy. Detected people are tracked independently in each camera view by means of a graph-based matching strategy, which estimates the best correspondences between consecutive people segmentations. Finally, 3D tracking and positioning of people is achieved by geometrical consistency analysis over the tracked 2D candidates, using head position (instead of object centroids) to increase robustness to foreground occlusions.

On the Mahalanobis Distance Classification Criterion for Multidimensional Normal Distributions

Many existing engineering works model the statistical characteristics of the entities under study as normal distributions. These models are eventually used for decision making, requiring in practice the definition of the classification region corresponding to the desired confidence level. Surprisingly enough, however, a great amount of computer vision works using multidimensional normal models leave unspecified or fail to establish correct confidence regions due to misconceptions on the features of Gaussian functions or to wrong analogies with the unidimensional case. The resulting regions incur in deviations that can be unacceptable in high-dimensional models. Here we provide a comprehensive derivation of the optimal confidence regions for multivariate normal distributions of arbitrary dimensionality. To this end, firstly we derive the condition for region optimality of general continuous multidimensional distributions, and then we apply it to the widespread case of the normal probability density function. The obtained results are used to analyze the confidence error incurred by previous works related to vision research, showing that deviations caused by wrong regions may turn into unacceptable as dimensionality increases. To support the theoretical analysis, a quantitative example in the context of moving object detection by means of background modeling is given.

Robust multi-camera 3D tracking from mono-camera 2d tracking using Bayesian Association

Visual tracking is essential for automatic scene understanding and surveillance of areas of interest. Monocular 2D tracking has been largely studied in the literature, but it usually provides inadequate or incomplete information for event interpretation. In addition, it proves insufficiently robust, due to view-point limitations and lack of depth information. However, the association of multiple cameras with overlapped fields of view allows the inference of 3D information and, thus, a richer description of the monitored scene.

Camera Localization UsingTrajectories and Maps

We propose a new Bayesian framework for automatically determining the position (location and orientation) of an uncalibrated camera using the observations of moving objects and a schematic map of the passable areas of the environment. Our approach takes advantage of static and dynamic information on the scene structures through prior probability distributions for object dynamics. The proposed approach restricts plausible positions where the sensor can be located while taking into account the inherent ambiguity of the given setting. The proposed framework samples from the posterior probability distribution for the camera position via data driven MCMC, guided by an initial geometric analysis that restricts the search space. A Kullback-Leibler divergence analysis is then used that yields the final camera position estimate, while explicitly isolating ambiguous settings. The proposed approach is evaluated in synthetic and real environments, showing its satisfactory performance in both ambiguous and unambiguous settings.

Adaptable Bayesian classifier for spatiotemporal nonparametric moving object detection strategies

Electronic devices endowed with camera platforms require new and powerful machine vision applications, which commonly include moving object detection strategies. To obtain high-quality results, the most recent strategies estimate nonparametrically background and foreground models and combine them by means of a Bayesian classifier. However, typical classifiers are limited by the use of constant prior values and they do not allow the inclusion of additional spatiodependent prior information. In this Letter, we propose an alternative Bayesian classifier that, unlike those reported before, allows the use of additional prior information obtained from any source and depending on the spatial position of each pixel.

Kernel bandwidth estimation for moving object detection in non-stabilized cameras

Sophisticated strategies have been recently proposed for the detection of moving objects in non-stabilized camera setups. These strategies model both, background and foreground, using spatio-temporal non-parametric estimation. However, as no appropriate methods for dynamical kernel bandwidth are available, high-quality results cannot be obtained in all situations. Here, an automatic and efficient kernel bandwidth estimation strategy for spatio-temporal modeling is proposed. Background kernel bandwidth is estimated via a novel statistical analysis of spatially weighted data distributions, whereas foreground kernel bandwidth is estimated using a mean shift based analysis of previously detected foreground regions.

Simultaneous 3D object tracking and camera parameter estimation by Bayesian methods and transdimensional MCMC sampling

Multi-camera 3D tracking systems with overlapping cameras represent a powerful mean for scene analysis, as they potentially allow greater robustness than monocular systems and provide useful 3D information about object location and movement. However, their performance relies on accurately calibrated camera networks, which is not a realistic assumption in real surveillance environments. Here, we introduce a multi-camera system for tracking the 3D position of a varying number of objects and simultaneously refining the calibration of the network of overlapping cameras. Therefore, we introduce a Bayesian framework that combines Particle Filtering for tracking with recursive Bayesian estimation methods by means of adapted transdimensional MCMC sampling. Additionally, the system has been designed to work on simple motion detection masks, making it suitable for camera networks with low transmission capabilities. Tests show that our approach allows a successful performance even when starting from clearly inaccurate camera calibrations, which would ruin conventional approaches.

Capabilities and limitations of mono-camera pedestrian-based autocalibration

Many environments lack enough architectural information to allow an autocalibration based on features extracted from the scene structure. Nevertheless, the observation over time of walking people can generally be used to estimate the vertical vanishing point and the horizon line in the acquired image. However, this information is not enough to allow the calibration of a general camera without presuming excessive simplifications. This paper presents a study on the capabilities and limitations of the mono-camera calibration methods based solely on the knowledge of the vertical vanishing point and the horizon line in the image. The mathematical analysis sets the conditions to assure the feasibility of the mono-camera pedestrian-based autocalibration. In addition, examples of applications are presented and discussed.

3D Tracking Using Multi-view Based Particle Filters

Visual surveillance and monitoring of indoor environments using multiple cameras has become a field of great activity in computer vision. Usual 3D tracking and positioning systems rely on several independent 2D tracking modules applied over individual camera streams, fused using geometrical relationships across cameras. As 2D tracking systems suffer inherent difficulties due to point of view limitations (perceptually similar foreground and background regions causing fragmentation of moving objects, occlusions), 3D tracking based on partially erroneous 2D tracks are likely to fail when handling multiple-people interaction. To overcome this problem, this paper proposes a Bayesian framework for combining 2D low-level cues from multiple cameras directly into the 3D world through 3D Particle Filters. This method allows to estimate the probability of a certain volume being occupied by a moving object, and thus to segment and track multiple people across the monitored area. The proposed method is developed on the basis of simple, binary 2D moving region segmentation on each camera, considered as different state observations. In addition, the method is proved well suited for integrating additional 2D low-level cues to increase system robustness to occlusions: in this line, a naive color-based (HSI) appearance model has been integrated, resulting in clear performance improvements when dealing with complex scenarios.

Unsupervised high-quality soccer field segmentation

Field segmentation is a fundamental step in many soccer applications. However, despite its importance, the existing segmentation algorithms are not able to provide successful results in complex scenarios. Moreover, they require the manual selection of several parameters, hindering their usability. Here, an unsupervised field segmentation strategy based on the estimation of the probability density function of the green chromacity of the image is proposed. Results show its ability to provide high-quality results in a wide variety of scenarios.