2,397 research outputs found
A Prototype Application for Long-time Behavior Modeling and Abnormal Events Detection
No full textIn this work we present a prototype application for modelling common behaviours from long-time observations of a scene. The core of the system is based on the method proposed in (Noceti and Odone, 2012), an adaptive technique for profiling patterns of activities on temporal data - coupling a string-based representation and an unsupervised learning strategy - and detecting anomalies - i.e., dynamic events diverging with respect to the usual dynamics. We propose an engineered framework where the method is adopted to perform an online analysis over very long time intervals (weeks of activity). The behaviour models are updated to accommodate new patterns and cope with the physiological scene variations. We provide a thorough experimental assessment, to show the robustness of the application in capturing the evolution of the scene dynamics
A Spectral Graph Kernel and Its Application to Collective Activities Classification
No full textIn this work we consider a machine learning setting where data are represented as graphs. First, we derive a kernel function which evaluates the similarity between graphs, while capturing pair-wise constraints between graph nodes. Second, we apply it to the problem of classifying collective activities: on this respect we first represent groups of people located in a spatial neighborhood as graphs, and then train a multi-class classifier able to capture the behavior of the groups. We evaluate our approach on a benchmark dataset and report a comparative analysis with other state-of-art methods which highlights the benefits of our approach
Semi-supervised learning of sparse representations to recognize people spatial orientation
No full textIn this paper we consider the problem of classifying people spatial orientation with respect to the camera viewpoint from 2D images. Structured multi-class feature selection allows us to control the amount of redundancy of our input data, while semi-supervised learning helps us coping with the intrinsic ambiguity of output labels. We model the multi-class classification problem with an all-pairs strategy based on the use of a coding matrix. A thorough experimental evaluation on the TUD Multiview Pedestrian benchmark dataset demonstrates the superiority of our approach w.r.t. state-of-the-art
Efficient pedestrian detection with group lasso
No full textIn this paper we deal with pedestrian detection and propose the use of group lasso to learn from data a compact and meaningful representation out of a high dimensional dictionary of local features. Group lasso, a regularized method with a sparsity-enforcing penalty term, has the very nice property of performing feature selection while preserving the internal structure of the dictionary. In our study we consider in particular variable-size HoGs, whose internal structure is composed by cells and blocks: since the entries of a block need to be computed together, the feature selection process is designed so to keep them or discard them all. The detection algorithm we obtain is a very neat procedure, simple to train and computationally efficient, which allows us to achieve a very good compromise between performance and computational cost, making the method very appropriate for video surveillance applications
Humans in groups: The importance of contextual information for understanding collective activities
No full textIn this work we consider the problem of modeling and recognizing collective activities performed by groups of people sharing a common purpose. For this aim we take into account the social contextual information of each person, in terms of the relative orientation and spatial distribution of people groups. We propose a method able to process a video stream and, at each time instant, associate a collective activity with each individual in the scene, by representing the individual – or target – as a part of a group of nearby people – the target group. To generalize with respect to the viewpoint we associate each target with a reference frame based on his spatial orientation, which we estimate automatically by semi-supervised learning. Then, we model the social context of a target by organizing a set of instantaneous descriptors, capturing the essence of mutual positions and orientations within the target group, in a graph structure. Classification of collective activities is achieved with a multi-class SVM endowed with a novel kernel function for graphs. We report an extensive experimental analysis on benchmark datasets that validates the proposed solution and shows significant improvements with respect to state-of-art results
Portable and fast text detection
No full textIn this paper, we describe an efficient pipeline for real-time text detection to be implemented on different architectures, with particular reference to smart phones. The text detection pipeline is based on a rather standard segmentation followed by a classification of each segmented connected component. Segmentation is performed by a linear implementation of MSER, state-of-the-art for text detection, where we control the overall computational cost of the method by computing a set of descriptive features as segmentation goes on. Classification is carried out by a cascade of SVM classifiers, where each layer captures different levels of complexity by means of an appropriate choice of descriptive features and kernel functions. Each detected text element, or character, is finally merged into lines of text and words. Further on, each element can be fed to a multi-class classifier that performs character recognition—this functionality is currently under development. We report experiments aiming at assessing the appropriateness of the text detection procedure, in terms of both performance and speed, when running on both x86 and ARM processors
Learning common behaviors from large sets of unlabeled temporal series
No full textThis paper is about extracting knowledge from large sets of videos, with a particular reference to the video-surveillance application domain. We consider an unsupervised framework and address the specific problem of modeling common behaviors from long-term collection of instantaneous observations. Specifically, such data describe dynamic events and may be represented as time series in an appropriate space of features. Starting off from a set of data meaningful of the common events in a given scenario, the pipeline we propose includes a data abstraction level, that allows us to process different data in a homogeneous way, and a behavior modeling level, based on spectral clustering. At the end of the pipeline we obtain a model of the behaviors which are more frequent in the observed scene, represented by a prototypical behavior, which we call a cluster candidate. We report a detailed experimental evaluation referring to both benchmark datasets and on a complex set of data collected in-house. The experiments show that our method compares very favorably with other approaches from the recent literature. In particular the results we obtain prove that our method is able to capture meaningful information and discard noisy one from very heterogeneous datasets with different levels of prior information available
“Hands On” Visual Recognition for Visually Impaired Users
No full textBlind or visually impaired (BVI) individuals are capable of identifying an object in their hands by combining the available visual cues (if available) with manipulation. It is harder for them to associate the object with a specific brand, a model, or a type. Starting from this observation, we propose a collaborative system designed to deliver visual feedback automatically and to help the user filling this semantic gap. Our visual recognition module is implemented by means of an image retrieval procedure that provides real-time feedback, performs the computation locally on the device, and is scalable to new categories and instances. We carry out a thorough experimental analysis of the visual recognition module, which includes a comparative analysis with the state of the art. We also present two different system implementations that we test with the help of BVI users to evaluate the technical soundness, the usability, and the effectiveness of the proposed concept.</jats:p
Mean BoF per quadrant: Simple and effective way to embed spatial information in bag of features
No full textThis paper proposes a new approach for embedding spatial information into a Bag of Features image descriptor, primarily meant for image retrieval. The method is conceptually related to Spatial Pyramids but instead of requiring fixed and arbitrary sub-regions where to compute region-based BoF, it relies on an adaptive procedure based on multiple partitioning of the image in four quadrants (the NE, NW, SE, SW regions of the image). To obtain a compact and efficient description, all BoF related to the same quadrant are averaged, obtaining four descriptors which capture the dominant structures of the main areas of the image, and then concatenated. The computational cost of the method is the same as BoF and the size of the descriptor comparable to BoF, but the amount of spatial information retained is considerable, as shown in the experimental analysis carried out on benchmarks
- …
