Gianluca Antonini

Counting Pedestrians in Video Sequences Using Trajectory Clustering

In this paper, we propose the use of lustering methods for automatic counting of pedestrians in video sequences. As input, we consider the output of those detection/tracking systems that overestimate the number of targets. Clustering techniques are applied to the resulting trajectories in order to reduce the bias between the number of tracks and the real number of targets. The main hypothesis is that those trajectories belonging to the same human body are more similar than trajectories belonging to different individuals. Several data representations and different distance/similarity measures are proposed and compared, under a common hierarchical clustering framework, and both quantitative and qualitative results are presented

Behavioral Priors for Detection and Tracking of Pedestrians in Video Sequences

In this paper we address the problem of detection and tracking of pedestrians in complex scenarios. The inclusion of prior knowledge is more and more crucial in scene analysis to guarantee flexibility and robustness, necessary to have reliability in complex scenes. We aim to combine image processing methods with behavioral models of pedestrian dynamics, calibrated on real data. We introduce Discrete Choice Models (DCM) for pedestrian behavior and we discuss their integration in a detection and tracking context. The obtained results show how it is possible to combine both methodologies to improve the performances of such systems in complex sequences.

Subagging for credit scoring models

The logistic regression framework has been for long time the most used statistical method when assessing customer credit risk. Recently, a more pragmatic approach has been adopted, where the first issue is credit risk prediction, instead of explanation. In this context, several classification techniques have been shown to perform well on credit scoring, such as support vector machines among others. While the investigation of better classifiers is an important research topic, the specific methodology chosen in real world applications has to deal with the challenges arising from the real world data collected in the industry. Such data are often highly unbalanced, part of the information can be missing and some common hypotheses, such as the i.i.d. one, can be violated. In this paper we present a case study based on a sample of IBM Italian customers, which presents all the challenges mentioned above. The main objective is to build and validate robust models, able to handle missing information, class unbalancedness and non-iid data points. We define a missing data imputation method and propose the use of an ensemble classification technique, subagging, particularly suitable for highly unbalanced data, such as credit scoring data. Both the imputation and subagging steps are embedded in a customized cross-validation loop, which handles dependencies between different credit requests. The methodology has been applied using several classifiers (kernel support vector machines, nearest neighbors, decision trees, Adaboost) and their subagged versions. The use of subagging improves the performance of the base classifier and we will show that subagging decision trees achieve better performance, still keeping the model simple and reasonably interpretable.

Multi-Layer Hierarchical Clustering of Pedestrian Trajectories for Automatic Counting of People in Video Sequences

In this paper we propose an approach to count the number of pedestrians, given a trajectory data set provided by a tracking system. The tracking process itself is treated as a black box providing us the input data. The idea is to apply a hierarchical clustering algorithm, using different data representations and distance measures, as a post-processing step. The final goal is to reduce the difference between the number of tracked pedestrians and the real number of individuals present in the scene.

Independent component analysis and support vector machine for face feature extraction

We propose Independent Component Analysis representation and Support Vector Machine classification to extract facial features in a face detection/localization context. The goal is to find a better space where project the data in order to build ten different face-feature classifiers that are robust to illumination variations and bad environment conditions. The method was tested on the BANCA database, in different scenarios: controlled conditions, degraded conditions and adverse conditions.

Modelling human perception of static facial expressions

A recent internet based survey of over 35,000 samples has shown that when different human observers are asked to assign labels to static human facial expressions, different individuals categorize differently the same image. This fact results in a lack of an unique ground-truth, an assumption held by the large majority of existing models for classification. This is especially true for highly ambiguous expressions, especially in the lack of a dynamic context. In this paper we propose to address this shortcoming by the use of discrete choice models (DCM) to describe the choice a human observer is faced to when assigning labels to static facial expressions. Different models of increasing complexity are specified to capture the causal effect between features of an image and its associated expression, using several combinations of different measurements. The sets of measurements we used are largely inspired by FACS but also introduce some new ideas, specific to a static framework. These models are calibrated using maximum likelihood techniques and they are compared with each other using a likelihood ratio test, in order to test for significance in the improvement resulting from adding supplemental features. Through a cross-validation procedure we assess the validity of our approach against overfitting and we provide a comparison with an alternative model based on Neural Networks for benchmark purposes.

Bayesian integration of a discrete choice pedestrian behavioral model and image correlation techniques for automatic multiobject tracking

In this paper we deal with the multiobject tracking problem in the particular case of pedestrians, assuming the detection step already done. We use a Bayesian framework to combine the likelihood term provided by an image correlation algorithm with a prior distribution given by a discrete choice model for pedestrian behavior, calibrated on real data. We aim to show how the combination of the image information with a model of pedestrian behavior can provides appreciable results in real and complex scenarios.

Robust infants face tracking using active appearance models: a mixed-state condensation approach

In this paper a new extension of the CONDENSATION algorithm, with application to infants face tracking, will be introduced. In this work we address the problem of tracking a face and its features in baby video sequences. A mixed state particle filtering scheme is proposed, where the distribution of observations is derived from an active appearance model. The mixed state approach combines several dynamic models in order to account for different occlusion situations. Experiments on real video show that the proposed approach augments the tracker robustness to occlusions while maintaining the computational time competitive.

A Discrete choice framework for acceleration and direction change behaviors in walking pedestrians

The walking process is interpreted as a sequence of decisions about where to put the next step. A dynamic and individual-based spatial discretization is used to represent the physical space. A behavioral framework for pedestrian dynamics based on discrete choice models is given. Direction change behaviors and acceleration behaviors are taken into account, both in a constrained and unconstrained formulation. The unconstrained direction changes (keep direction, toward destination) and acceleration (free flow acceleration) behaviors are the same as those introduced in our previous work. In this paper we focus on the definition of the constrained counterparts. A leader follower behavior is interpreted as a constrained acceleration while collision avoidance behavior as a constrained direction change. The spatial correlation structure in the choice set deriving from a simultaneous choice of speed regimes and radial directions is taken into account specifying a cross nested logit model (CNL). Quantitative results are presented, obtained by maximum likelihood estimation on a real data set with more than 10 thousands observed positions, manually tracked from video sequences.

Mining actionable partial orders in collections of sequences

Mining frequent partial orders from a collection of sequences was introduced as an alternative to mining frequent sequential patterns in order to provide a more compact/understandable representation. The motivation was that a single partial order can represent the same ordering information between items in the collection as a set of sequential patterns (set of totally ordered sets of items). However, in practice, a discovered set of frequent partial orders is still too large for an effective usage. We address this problem by proposing a method for ranking partial orders with respect to significance that extends our previous work on ranking sequential patterns. In experiments, conducted on a collection of visits to a website of a multinational technology and consulting firm we show the applicability of our framework to discover partial orders of frequently visited webpages that can be actionable in optimizing effectiveness of web-based marketing.