Evidential reasoning framework for object tracking

2005

Pose estimation involves reconstructing the configuration of a moving body from images sequences. In this paper we present a general framework for pose estimation of unknown objects based on Shafer’s evidential reasoning. During learning an evidential model of the object is built, integrating different image features to improve both estimation robustness and precision. All the measurements coming from one or more views are expressed as belief functions, and combined through Dempster’s rule. The best pose estimate at each time step is then extracted from the resulting belief function by probabilistic approximation. The choice of a sufficiently dense training set is a critical problem. Experimental results concerning a human tracking system are shown.

In example-based pose estimation, the configuration or “pose” of an evolving object is sought given visual evidence, having to rely uniquely on a set of examples. We assume here that, in a training stage, a number of feature measurements is extracted from the available images, while an “oracle” provides us with the true object pose at each instant. In this scenario, a sensible approach consists in learning maps from features to poses, using the information provided by the training set. In particular, multivalued mappings linking feature values to set of training poses can be easily constructed. A probability measure on any feature space is then naturally mapped to a convex set of probabilities on the set of training poses, in a form of a “belief function”. Given a test image, its feature measurements translate into a collection of belief functions on the set of training poses, which when combined yield there an entire family of probability distributions. From the latter either a single central pose estimate or a set of extremal estimates can be computed, together with a measure of how reliable the estimate is. We call this technique “Belief Modeling Regression” (BMR). We illustrate BMR’s performance in an application to human pose recovery, showing how it outperforms our implementation of both Relevant Vector Machine and Gaussian Process Regression. We discuss motivation and advantages of the proposed approach with respect to its competitors and outline an extension of this technique to tracking.

Proc. of MTNS2000, 2000

In this position paper we propose the use of the Distributional Clauses Particle Filter in conjunction with a model-based 3D object tracking method in monocular camera sequences. We describe the model based object tracking method that is based on contour and edge features for 3D pose relative estimation. We also describe the application of the Distributional Clauses Particle Filter that takes into account inputs from object tracking. We argue that objects' dynamics can be modeled via probabilistic rules, which makes possible to predict and utilise a pose hypothesis space for fully occluded or 'invisible' (hidden-away) objects that may re-appear in the camera field of view. Important issues, such as losing track of the object in a 'total occlusion' scenario, are discussed.

In example-based pose estimation, the configuration or “pose” of an evolving object is sought given visual evidence, having to rely uniquely on a set of examples. In a training stage, a number of features are extracted from the available images and ground truth poses are acquired. In this scenario, a sensible approach consists in learning maps from features to poses, using the training data. In particular, multi-valued mappings linking feature values to a set of training poses can be easily constructed: in this paper, we propose to use these mappings for pose estimation. In the proposed method, a probability measure on any feature space is naturally mapped to a convex set of probabilities on the set of training poses, in a form of a “belief function”. Given a test image, its features translate into a collection of belief functions on the set of training poses, which when combined yield there an entire family of probability distributions. From the latter, both a single, central pose estimate and a set of extremal estimates can be computed, together with a measure of how reliable the estimate is. We call this technique “Belief Modeling Regression”. We demonstrate its effectiveness by comparing it to popular mapping techniques such as Gaussian Process and Relevance Vector Regression under a varied set of experiments.

International Journal of Approximate Reasoning, 2009

This work proposes a novel filtering algorithm that constitutes an extension of Bayesian particle filters to the Dempster-Shafer theory. Our proposal solves the multi-target problem by combining evidences from multiple heterogeneous and unreliable sensors. The modelling of uncertainty and absence of knowledge in our approach is specially attractive since it does not require to specify prior nor conditionals that might be difficult to obtain in complex problems.

Lecture Notes in Computer Science, 2007

We present a probabilistic framework for component-based automatic detection and tracking of objects in video. We represent objects as spatio-temporal two-layer graphical models, where each node corresponds to an object or component of an object at a given time, and the edges correspond to learned spatial and temporal constraints. Object detection and tracking is formulated as inference over a directed loopy graph, and is solved with non-parametric belief propagation. This type of object model allows object-detection to make use of temporal consistency (over an arbitrarily sized temporal window), and facilitates robust tracking of the object. The two layer structure of the graphical model allows inference over the entire object as well as individual components. AdaBoost detectors are used to define the likelihood and form proposal distributions for components. Proposal distributions provide 'bottomup' information that is incorporated into the inference process, enabling automatic object detection and tracking. We illustrate our method by detecting and tracking two classes of objects, vehicles and pedestrians, in video sequences collected using a single grayscale uncalibrated carmounted moving camera.

The Fourth International Conference onComputer and Information Technology, 2004. CIT '04.

We present a novel method for tracking the motion of an articulated structure in a video sequence. The analysis of articulated motion is challenging because of the potentially large number of degrees of freedom (DOFs) of an articulated body. For particle filter based algorithms, the number of samples required with high dimensional problems can be computationally prohibitive. To alleviate this problem, we represent the articulated object as an undirected graphical model (or Markov Random Field, MRF) in which soft constraints between adjacent subparts are captured by conditional probability distributions. The graphical model is extended across time frames to implement a tracker. The tracking algorithm can be interpreted as a belief inference procedure on a dynamic Bayesian network. The discretisation of the state vectors makes it possible to utilise the efficient belief propagation (BP) and mean field (MF) algorithms to reason in this network. Experiments on real video sequences demonstrate that the proposed method is computationally efficient and performs well in tracking the human body.

In this paper we propose a novel real-time tracking algorithm robust with respect to several common errors occurring in object detection systems, especially in the presence of total or partial occlusions. The algorithm takes into account the history of each object, whereas most other methods base their decisions on only the last few frames. More precisely, it associates each object with a state encoding the relevant information of its past history, that enable the most appropriate way of assigning an identity to the object on the basis of its current and past conditions. Thus, strategies that are more complex but also riskier are only applied when the algorithm is confident that is appropriate to do so. An experimental evaluation of the algorithm has been performed using the PETS2010 database, comparing the obtained performance with the results of the PETS 2010 contest participants.

Image and Vision Computing, 2006

A motion detection and tracking algorithm for human and car activity surveillance is presented and evaluated by using the Pets'2000 test sequence. Proposed approach uses a temporal fusion strategy by using the history of events in order to improve instantaneous decisions. Normalized indicators updated at each frame summarize history of specific events. For the motion detection stage a fast updating algorithm of the background reference is proposed. The control of the updating at each pixel is based on a stability indicator estimated from inter-frame variations. The tracking algorithm uses a determinist region based approach. A belief indicator representing the tracking consistency for each object allows to solve defined ambiguities at the tracking level. A second specific tracking indicator representing the identity quality of each tracked object is updated by integrating objects interaction. Tracking indicators permit to propagate uncertainties on higher levels of the interpretation and also are used directly in the tracking performance evaluation.