Catálogo de publicaciones - libros

The human activity recognition in video sequences is a field where many types of classifiers have been used as well as a wide range of input features that feed these classifiers. This work has a double goal. First of all, we extracted the most relevant features for the activity recognition by only utilizing motion features provided by a simple tracker based on the 2D centroid coordinates and the height and width of each person’s blob. Second, we present a performance comparison among seven different classifiers (two Hidden Markov Models (HMM), a J.48 tree, two Bayesian classifiers, a classifier based on rules and a Neuro-Fuzzy system). The video sequences under study present four human activities (inactive, active, walking and running) that have been manual labeled previously. The results show that the classifiers reveal different performance according to the number of features employed and the set of classes to sort. Moreover, the basic motion features are not enough to have a complete description of the problem and obtain a good classification.

A special kind of surveillance problem is the monitoring of wide enclosed areas with difficult access and changing environments. The characteristics of this kind of problem recommend a solution based on a multi-agent system, where several robots cooperate to solve the problem. A multi-robot system for surveillance in these kinds of environments has been designed and simulated on the Gazebo 3D simulator. Typical surveillance tasks are simulated and experimental results are shown.

A key aspect in visual surveillance systems is robust movement segmentation, which is still a difficult and unresolved problem. In this paper, we propose an architecture based on a two-layer image-processing modules: General Tracking Layer (GTL) and Context Layer (CL). GTL describe a generic multipurpose tracking process for video-surveillance systems. CL is designed as a symbolic reasoning system that manages the symbolic interface data between GTL modules in order to asses a specific situation and take the appropriate decision about visual data association. Our architecture has been used to improve the association process of a tracking system and tested in two different scenarios to show the advantages in improved performance and output continuity.

One of the main characteristics of a video surveillance system is its reliability. To this end, it is needed that the images captured by the videocameras are an accurate representation of the scene. Unfortunately, some activities can make the proper operation of the cameras fail, distorting in some way the images which are going to be processed. When these activities are voluntary, they are usually called sabotage, which include partial o total occlusion of the lens, image defocus or change of the field of view.

In this paper, we will analyze the different kinds of sabotage that could be done to a video surveillance system, and some algorithms to detect these inconveniences will be developed. The experimental results show good performance in the detection of sabotage situations, while keeping a very low false alarm probability.

Identification of aircrafts by means of radar when no cooperation exists (Non-Cooperative Target Identification, NCTI) tends to be based on simulations. To improve them, and hence the probability of correct identification, right values of permittivity and permeability need to be used. This paper describes a method for the estimation of the electromagnetic properties of materials as a part of the NCTI problem. Different heuristic optimization algorithms such as Genetic Algorithms (GA) and Particle Swarm Optimization (PSO), as well as other approaches like Artificial Neural Networks (ANN), are applied to the reflection coefficient obtained via free-space measurements in an anechoic chamber. Prior to the comparison with real samples, artificial synthetic materials are generated to test the performance of these bio-inspired algorithms.

This work is placed in the context of computer vision and ubiquitous multimedia access. It deals with the development of an automated system for human behavior analysis and prediction using context features as a representative descriptor of human posture. In our proposed method, an action is composed of a series of features over time. Therefore, time sequential images expressing human action are transformed into a feature vector sequence. Then the feature is transformed into symbol sequence. For that purpose, we design a posture codebook, which contains representative features of each action type and define distances to measure similarity between feature vectors. The system is also able to predict next performed motion. This prediction helps to evaluate and choose current action to show.

This paper deals with the problem of estimating the following road sign parameters: height, dimensions, visibility distance and partial occlusions. This work belongs to a framework whose main applications involve road sign maintenance, driver assistance, and inventory systems. From this paper we suggest a multisensory system composed from two cameras, a GPS receiver, and a distance measurement device, all of them installed in a car. The process consists of several steps which include road sign detection, recognition and tracking , and road signs parameters estimation. From some trigonometric properties, and a camera model, the information provided by the tracking subsystem and the distance measurement sensors, we estimate the road signs parameters. Results show that the described calculation methodology offers a correct estimation for all types of traffic signs.

In this work an efficient and robust technique of data association will be developed as a search in the hypotheses space defined by the possible association between detections and tracks. The full data association problem in visual tracking is formulated as a hypotheses search with a heuristic evaluation function to take into account structural and specific information such as distance, shape, colour, etc. This heuristic should represent the real problem so that its optimization leads to the solution of each situation. In order to guarantee performance in real time, the search process will have assigned a bounded amount of time to give the solution. The number of evaluations is restricted to accomplish this bound. The use of Estimation Distribution Algorithms (EDA) allows the application of an Evolutionary Computation technique to search in the hypothesis space. The performance of alternative algorithms used to provide the solution with this time constraint will be compared considering complex situations.

One of the major forces for understanding forest fire risk and behavior is the fire fuel. Fire risk and behavior depend on the fuel properties such as moisture content. Context information on vegetation water content is vital for understanding the processes involved in initiation and propagation of forest fires. In that sense, a novel method was tested to estimate vegetation canopy water content (CWC) from simulated MODIS satellite data. An inversion of a radiative transfer model called Forest Light Interaction-Model (FLIM) from performed using evolutionary computation. CWC is critical, among other applications, in wildfire risk assessment since a decrease in CWC causes higher probability to have wildfire occurrence. Simulations were carried out with the FLIM model for a wide range of forest canopy characteristics and CWC values. A 50 subsample of the simulations was used for the training process and 50 for the validation providing a RMSE=0.74 and r2=0.62. Further research is needed to apply this method on real MODIS images.

Reinforcement Learning (RL) approaches are, very often, rendered useless by the statistics of the required sampling process. This paper shows how RL is essentially made possible by abandoning the state feedback during training episodes. The resulting new method, (FF learning), employs a return estimator for pairs of a state and a parameter vector. FF learning is particularly suitable for the learning of controllers, e.g. for robotics applications, and yields learning rates unprecedented in the RL context.

This paper introduces the method formally and proves a lower bound on its performance. Practical results are provided from applying FF learning to several scenarios based on the collision avoidance behavior of a mobile robot.