Human motion prediction from visual tracking is a challenging problem with a wide array of applications such as robotics, video surveillance and situation understanding. Recently, planning-based methods -which assume that people move by planning over a cost function- have emerged as one of the most promising alternatives. Nevertheless, state of the art planning based algorithms have shortcomings regarding their computational complexity and ability to predict for arbitrary time intervals. This paper addresses these shortcomings by leveraging alternative planning techniques (Fast Marching Method) and formulating efficient algorithms for goal estimation and full spatiotemporal prediction with lower complexity than comparable approaches. In preliminary experiments, the proposed method significantly outperforms the accuracy of the current state-of-the-art approach while reducing the computation time by a factor of 30 using a parallel version of our algorithm.

For mobile robots which operate in human populated environments, modeling social interactions is key to understand and reproduce people's behavior. A promising approach to this end is Inverse Reinforcement Learning (IRL) as it allows to model the factors that motivate people's actions instead of the actions themselves. A crucial design choice in IRL is the selection of features that encode the agent's context. In related work, features are typically chosen ad hoc without systematic evaluation of the alternatives and their actual impact on the robot's task. In this paper, we introduce a new software framework to systematically investigate the effect features and learning algorithms used in the literature. We also present results for the task of socially compliant robot navigation in crowds, evaluating two different IRL approaches and several feature sets in large-scale simulations. The results are benchmarked according to a proposed set of objective and subjective performance metrics.

In many computer vision related applications it is necessary to distinguish between the background of an image and the objects that are contained in it. This is a difficult problem because of the constraints imposed by the available time and the computational cost of robust object extraction algorithms. This report describes a new method that benefits from state of the art background/foreground classification combined with the strong theoretical foundations of clustering. The pixels on the scene background are modeled as Mixtures of Gaussians and the output of the classification process are continuous values representing the likelihood that each pixel belongs to the foreground. The clustering is based on a Self Organizing Network (SON) which has a robust initialization schema and is able to find the number of objects in an image or grid. The algorithm's complexity is linear with respect to the number of pixels or cells.

We present to the robotic community a fully autonomous navigation solution for mobile robots operating in urban pedestrian areas. We introduce our robots and the experimental zone, overview the architecture of the navigation framework, and present the results after 3.5km of autonomous navigation. We expose the main lessons learnt by the scientific team and identify the issues to improve future works.

Modeling and predicting human and vehicle motion is an active research domain. Due to the difficulty of modeling the various factors that determine motion (e.g., internal state and perception), this is often tackled by applying machine learning techniques to build a statistical model, using as input a collection of trajectories gathered through a sensor (e.g., camera and laser scanner), and then using that model to predict further motion. Unfortunately, most current techniques use offline learning algorithms, meaning that they are not able to learn new motion patterns once the learning stage has finished. In this paper, we present an approach where motion patterns can be learned incrementally and in parallel with prediction. Our work is based on a novel extension to hidden Markov models (HMMs) - called growing hidden Markov models - which gives us the ability to incrementally learn both the parameters and the structure of the model.

In many computer vision related applications it is necessary to distinguish between the background of an image and the objects that are contained in it. This is a difficult problem because of the double constraint on the available time and the computational cost of robust object extraction algorithms. This paper builds upon former work on combining the strong theoretical foundations of clustering with the speed of other approaches. It is based on a novel self organizing network (SON) which has a robust initialization schema and is able to find the number of objects in an image or grid. The main contribution of our extension is that it eliminates the use of a threshold, allowing the algorithm to work on continuous, while having a complexity that remains linear with respect to the number of pixels or cells.

Despite their popularity, occupancy grids cannot be directly applied to problems where the identity of the objects populating an environment needs to be taken into account (e.g., object tracking, scene interpretation, etc.), in this cases it is necessary to postprocess the grid in order to extract object information. This paper approaches the problem by proposing a novel algorithm inspired on image segmentation techniques. The proposed approach works without prior knowledge about the number of objects to be detected and, at the same time, is very fast. This is possible thanks to the use of a novel self organizing network (SON) coupled with a dynamic threshold. Our experimental results on both real and simulated data show that our approach is robust and able to operate at normal camera frame rate

Abstract — Image segmentation is a critical task in computer vision. In the context of motion detection, a very popular segmentation approach is background substraction which consists in classifiying the pixels as background and foreground. Then, the foreground pixels are grouped together to find objects, this task is known as object extraction. There are several different approaches to object extraction (eg connected component labeling, morphological operators, size thresholding and clustering) amongst them, cluster based approaches are, probably, the ones with a stronger theoretical foundation. However, their application to object extraction is difficult because of three problems: a) need to know the number of objects to be detected beforehand, b) high sensibility to initialization due to a trend to get stuck in local minima and c) high complexity which difficults their application in real-time. This paper proposes an algorithm which aims to combine the strong theoretical foundations of clustering with the speed of other approaches. This is possible due to the introduction of a novel Self Organizing Network (SON) which has a robust initialization schema and is able to find the number of clusters in the image. The algorithm has a time complexity of order NM where N is the number of foreground pixels in the image and M is the number of nodes in the SON. I.

Vehicle navigation in dynamic environments is an important challenge, especially when the motion of the objects populating the environment is unknown. Traditional motion planning approaches are too slow to be applied in real-time to this domain, hence, new techniques are needed. Recently, iterative planning has emerged as a promising approach. Nevertheless, existing iterative methods do not provide a way to estimate the future behavior of moving obstacles and use the resulting estimates in trajectory computation. This paper presents an iterative planning approach that addresses these two issues. It consists of two complementary methods: 1) a motion prediction method which learns typical behaviors of objects in a given environment; 2) an iterative motion planning technique based on the concept of velocity obstacles.

This paper proposes a technique to obtain long term estimates of the motion of a moving object in a structured environment. Objects moving in such environments often participate in typical motion patterns which can be observed consistently. Our technique learns those patterns by observing the environment and clustering the observed trajectories using any pairwise clustering algorithm. We have implemented our technique using both simulated and real data coming from a vision system. The results show that the technique is general, produces long-term predictions and is fast enough for its use in real time applications.

Vehicle navigation in dynamic environments is an important challenge, especially when the motion of the objects populating the environment is unknown. Traditional motion planning approaches are too slow to be applied in real-time to this domain, hence, new techniques are needed. Recently, iterative planning has emerged as a promising approach. Nevertheless, existing iterative methods do not provide a way to estimate the future behaviour of moving obstacles and use the resulting estimates in trajectory computation. This paper presents an iterative planning approach that addresses these two issues. It consists of two complementary methods: 1) a motion prediction method which learns typical behaviours of objects in a given environment; 2) an iterative motion planning technique based on the concept of velocity obstacles.