Alfredo Petrosino

A Self-Organizing Approach to Background Subtraction for Visual Surveillance Applications

Detection of moving objects in video streams is the first relevant step of information extraction in many computer vision applications. Aside from the intrinsic usefulness of being able to segment video streams into moving and background components, detecting moving objects provides a focus of attention for recognition, classification, and activity analysis, making these later steps more efficient. We propose an approach based on self organization through artificial neural networks, widely applied in human image processing systems and more generally in cognitive science. The proposed approach can handle scenes containing moving backgrounds, gradual illumination variations and camouflage, has no bootstrapping limitations, can include into the background model shadows cast by moving objects, and achieves robust detection for different types of videos taken with stationary cameras. We compare our method with other modeling techniques and report experimental results, both in terms of detection accuracy and in terms of processing speed, for color video sequences that represent typical situations critical for video surveillance systems.

The SOBS algorithm: What are the limits?

The Self-Organizing Background Subtraction (SOBS) algorithm implements an approach to moving object detection based on the neural background model automatically generated by a self-organizing method, without prior knowledge about the involved patterns. Such adaptive model can handle scenes containing moving backgrounds, gradual illumination variations and camouflage, can include into the background model shadows cast by moving objects, and achieves robust detection for different types of videos taken with stationary cameras. Moreover, the introduction of spatial coherence into the background update procedure leads to the so-called SC-SOBS algorithm, that provides further robustness against false detections. The paper includes extensive experimental results achieved by the SOBS and the SC-SOBS algorithms on the dataset made available for the Change Detection Challenge at the IEEE CVPR2012.

A fuzzy spatial coherence-based approach to background/foreground separation for moving object detection

The detection of moving objects from stationary cameras is usually approached by background subtraction, i.e. by constructing and maintaining an up-to-date model of the background and detecting moving objects as those that deviate from such a model. We adopt a previously proposed approach to background subtraction based on self-organization through artificial neural networks, that has been shown to well cope with several of the well known issues for background maintenance. Here, we propose a spatial coherence variant to such approach to enhance robustness against false detections and formulate a fuzzy model to deal with decision problems typically arising when crisp settings are involved. We show through experimental results and comparisons that higher accuracy values can be reached for color video sequences that represent typical situations critical for moving object detection.

Adjusted F-measure and kernel scaling for imbalanced data learning

Rare events are involved in many challenging real world classification problems, where the minority class is usually the most expensive to sample and to label. As a consequence, training data are often imbalanced, presenting an heavily skewed distribution of labels. Using conventional classification techniques produces biased results, as the classifier may easily show a very good performance on the over-represented class and a very poor performance on the under-represented class: the former dominates the learning process and tends to attract all predictions. Furthermore, the classical accuracy measure is misleading, as it assumes equal importance for the true positives and the true negatives. We propose a classification procedure based on Support Vector Machine able to effectively cope with data imbalance. Using a first step approximate solution and then a suitable kernel transformation, we enlarge asymmetrically space around the class boundary, compensating data skewness. We also propose an accuracy measure, named AGF, that properly accounts for the different misclassification costs of the two classes. Tests on real world data from a public repository show that the proposed approach outperforms its competitors.

Rough fuzzy set based scale space transforms and their use in image analysis

In this paper we present a multi-scale method based on the hybrid notion of rough fuzzy sets, coming from the combination of two models of uncertainty like vagueness by handling rough sets and coarseness by handling fuzzy sets. Marrying both notions lead to consider, as instance, approximation of sets by means of similarity relations or fuzzy partitions. The most important features are extracted from the scale spaces by unsupervised cluster analysis, to successfully tackle image processing tasks. Here, we report some results achieved by applying the method to multi-class image segmentation and edge detection, but it can be shown to be successfully applied to texture discrimination problem too.

Towards Benchmarking Scene Background Initialization

Given a set of images of a scene taken at different times, the availability of an initial background model that describes the scene without foreground objects is the prerequisite for a wide range of applications, ranging from video surveillance to computational photography. Even though several methods have been proposed for scene background initialization, the lack of a common groundtruthed dataset and of a common set of metrics makes it difficult to compare their performance. To move first steps towards an easy and fair comparison of these methods, we assembled a dataset of sequences frequently adopted for background initialization, selected or created ground truths for quantitative evaluation through a selected suite of metrics, and compared results obtained by some existing methods, making all the material publicly available.

The 3dSOBS+ algorithm for moving object detection

We propose the 3dSOBS+ algorithm, a newly designed approach for moving object detection based on a neural background model automatically generated by a self-organizing method. The algorithm is able to accurately handle scenes containing moving backgrounds, gradual illumination variations, and shadows cast by moving objects, and is robust against false detections for different types of videos taken with stationary cameras. Experimental results and comparisons conducted on the Background Models Challenge benchmark dataset demonstrate the improvements achieved by the proposed algorithm, that compares well with the state-of-the-art methods.

Stopped Object Detection by Learning Foreground Model in Videos

The automatic detection of objects that are abandoned or removed in a video scene is an interesting area of computer vision, with key applications in video surveillance. Forgotten or stolen luggage in train and airport stations and irregularly parked vehicles are examples that concern significant issues, such as the fight against terrorism and crime, and public safety. Both issues involve the basic task of detecting static regions in the scene. We address this problem by introducing a model-based framework to segment static foreground objects against moving foreground objects in single view sequences taken from stationary cameras. An image sequence model, obtained by learning in a self-organizing neural network image sequence variations, seen as trajectories of pixels in time, is adopted within the model-based framework. Experimental results on real video sequences and comparisons with existing approaches show the accuracy of the proposed stopped object detection approach.

Multi-feature adaptive classifiers for SAR image segmentation

Abstract We propose a multifeature scheme for terrain classification in SAR image analysis. Different neural classifiers, trained on different features of the same sample space, are combined by using a non-linear ensemble method. The feature extraction modules are chosen in order to discover the textural and contextual characteristics within the neighbourhood of each pixel. Comparisons with classical data fusion techniques and consensus schema are reported.

MATRIOSKA: A Multi-level Approach to Fast Tracking by Learning

In this paper we propose a novel framework for the detection and tracking in real-time of unknown object in a video stream. We decompose the problem into two separate modules: detection and learning. The detection module can use multiple keypoint-based methods (ORB, FREAK, BRISK, SIFT, SURF and more) inside a fallback model, to correctly localize the object frame by frame exploiting the strengths of each method. The learning module updates the object model, with a growing and pruning approach, to account for changes in its appearance and extracts negative samples to further improve the detector performance. To show the effectiveness of the proposed tracking-by-detection algorithm, we present quantitative results on a number of challenging sequences where the target object goes through changes of pose, scale and illumination.