Video anomaly detection (VAD) has been densely explored in the last few years, mostly in single-camera scenarios. Despite significant advancements in this field, effectiveness is still seriously compromised in challenging environments (e.g., varying lighting conditions, under partial occlusions, and in crowded environments). For the sake of affordable data annotation, the most relevant methods assume the weakly supervised paradigm, where the label is available only at the video level (WS-VAD). Also, these methods are conventionally designed for single-camera mode and do not consider the multi-view information yielded from overlapping surveillance cameras, which is very common in practical scenarios. In this work, we started by systematically evaluating the WS-VAD performance that can be attained when different camera combinations are used as data sources. Interestingly, we observed that the rule "the more cameras, the better" should not be assumed, as there were always particular subsets of cameras that consistently outperformed the remaining configurations. Upon these conclusions, we present a semi-automated procedure to identify the optimal camera sources based on the image features/characteristics (distance, pose, and lighting) each one is capturing. Extensive experiments were carried out in three overlapping multi-camera datasets, which suggest that 1) multi-camera schemes consistently outperform single-camera methods and - most interestingly - 2) the correlation between the data acquired by the different cameras severely impacts performance, turning the selection of cameras a crucial step in VAD. Our findings open an intriguing research topic about methods/algorithms that filter out/select the camera sources we should use in overlapping camera scenarios.

This paper describes and evaluates a Mean-Shift-based (MS) approach to an instance of the Cooperative Target Observation (CTO) problem domain. A performance comparison is presented with a k-means-based approach to the baseline implementation published to the CTO problem. Inspired by the idea of modeling the problem for urban centers in which the movement of targets is restricted to the streets and roads, we also evaluate the effect of the movement of the targets being restricted to a rectangular grid on the relative performance of the algorithms. We conclude that the MS-based approach is superior to the k-means-based approach and that the target motion restricted to a grid improves both algorithms' performance but does not change its relative positions.

This paper proposes to segment face images into six classes (eyes, nose, mouth, hair/eyebrows/beard, skin, and background) by classifying pixels based on the texture features calculated in a neighborhood of each pixel. Leung-Malik filter banks are applied to the color images for feature extraction and Random Projections are used to reduce data dimensionality. In order to perform pixel classification, manually labeled images are used to train a Multi-Quadric Radial Basis Function Neural Network, with centers selected by the Fast Condensed Nearest Neighbor algorithm. Quantitative and qualitative results are presented and demonstrate that the methodology can correctly segment most of the class labels with high effectiveness rate, comparable with the results achieved by state-of-art methods.

Traffic classification by application class provides useful information for various tasks of network engineering and administration. However, offline classification of flows has limited its practical application to auditing tasks, long-term planning and other analytical issues. Therefore, research on traffic classification now moves towards the search for accurate and efficient methods of classification in order to meet online tasks such as traffic monitoring and shaping and other specific-application operations. In this work we apply the One-Against-All Approach (OAA) for two online classification strategies based on statistical features of TCP sub-flows. One uses the first N packets of the bi-directional TCP session and the other applies to sub-flows of the N packets starting at a random position in the flow. In our variant of the OAA approach, the problem of classifying an object in one of M classes is reduced to M binary classification problems with an associated decision rule, with each of them possibly using a different subset of features and sub-flow size. We investigated the effect of variation in the amount of N on the results of classification and the smaller set of variables in each of the above problems. This study used the Naïve Bayes classifier.

This paper examines the performance of a new Hidden Markov Model (HMM) structure used as the core of an Internet traffic classsifier and compares the results against other models present in the literature. Traffic modeling and classification find importance in many areas such as bandwidth management, traffic analysis, prediction and engineering, network planning, Quality of Service provisioning and anomalous traffic detection. The new HMM structure, which takes into account the packet payload size (PS) and the inter-packet times (IPT) sequences, is obtained by concatenation of a first part which is framed with a HMM profile with another part whose structure is that of a fully-connected HMM. The first part captures the specific properties of the initial protocol packets while the second part captures the statistical properties of the whole sequence present in the flow. Models generated are found to increase the accurate in classifying different traffic classes in the analysed dataset. The average accuracy obtained by the classifier is 62.5% having seen only five packets, 80.0% after examining 13 packets and 95.5% after seeing the unidirectional entire flow.

The growing demand for link bandwidth and node capacity is a frequent phenomenon in IP network backbones. Within this context, traffic prediction is essential for the network operator. Traffic prediction can be undertaken based on link traffic or on origin-destination (OD) traffic which presents better results. This work investigates a methodology for traffic prediction based on multidimensional OD traffic, focusing on the stage of short-term traffic prediction using Principal Components Analysis as a technique for dimensionality reduction and a Local Linear Model based on K-means as a technique for prediction and trend analysis. The results validated with data on a real network present a satisfactory margin of error for use in practical situations.

This paper examines the formation of self-organizing feature maps (SOFM) by the direct optimization of a cost function through a genetic algorithm (GA). The resulting SOFM is expected to produce simultaneously a topologically correct mapping between input and output spaces and a low quantization error. The proposed approach adopts a cost (fitness) function which is a weighted combination of indices that measure these two aspects of the map quality, specifically, the quantization error and the Pearson correlation coefficient between the corresponding distances in input and output spaces. The resulting maps are compared with those generated by the Kohonen's self-organizing map (SOM) algorithm in terms of the quantization error (QE), the weighted topological error (WTE) and the Pearson correlation coefficient (PCC) indices. The experiments show the proposed approach produces better values of the quality indices as well as is more robust to outliers.

This work presents an Internet traffic classification methodology based on statistical discriminators and cluster analysis. An accuracy identification of Internet applications is an important research area, because it is directly related to solve many network problems such as: quality of service (QoS), traffic control, security, network management and operation. The main difference to previous approaches lies in the discriminators use; rather than using only one set of discriminators for all classes we use a set of different statistical discriminators for each traffic class. Using real traces into the training and classification phases, we validated the methodology for P2P traffic class.