Section: New Results

Ambiant Intelligence And Robotic Systems

Adaptation of autonomous vehicle traffic to perturbations

Participants : Mohamed Tlig, Olivier Buffet.

Olivier Simonin, a former member of the MAIA team, is an external collaborator from INSA-Lyon.

The aim of the European project InTraDE is to propose more efficient ways to handle containers in seaports through the use of IAVs (Intelligent Autonomous Vehicles).

In his PhD thesis, Mohamed Tlig considers the displacements of numerous such IAVs whose routes are a priori planned by a supervisor. However, in such a large and complex system, different unexpected events can arise and degrade the traffic: failure of a vehicle, human mistake while driving, obstacle on roads, local re-planning, and so on.

In 2013, we have started looking at improving vehicle flows in complete road networks. In particular, we have proposed an approach that allows multiple flows of vehicles to cross an intersection without stopping, allowing to reduce delays as well as energy consumption. This has led to a publication in ICALT-14 [30] , with more details in a research report [41] .

This year, we have made a further step by coordinating the controller agents located in each of the network's intersections. More precisely, they are constrained to let the vehicles alternate at the same frequency —at the expense of potentially reducing the maximum flow of some roads— and a distributed algorithm offsets these “signals” so as to optimize either the energy consumption, or the time spent in the network. This tends to induce “green waves” wherever possible, i.e., to prevent vehicles from having to slow down before a traffic light. This work has been presented at ECAI-14 [31] .

Platooning: safe and precise virtual hooking mechanism or automated vehicles

Participants : Jano Yazbeck, Alexis Scheuer, François Charpillet.

Among the several goals that we were trying to achieve in InTraDE, we were interested in platooning too. In her PhD thesis, Jano Yazbeck considers Platooning as a technique that aims at steering , safely and precisely, a train of vehicles along a path generated by a leader which can be driven by a human. Thus the trajectory is unknown to the followers. Platooning is considered in this project in order to move containers efficiently from the discharge zones of ships to the storage areas.

To obtain a safe and precise platooning, we aim at controlling the longitudinal and lateral behaviors of each vehicle of the platooning. On the one hand, the longitudinal controller computes a longitudinal velocity (or acceleration) which avoids collisions between vehicles by maintaining a safe inter-distance between each couple of successive vehicles. On the other hand, the lateral controller computes an angular velocity or a steering angle so that the vehicle follows precisely the leader's path. These two controllers can be decoupled and computed separately when the convoy moves at a low velocity.

This year, we proposed a platooning algorithm based on a near-to-near decentralized approach which has been published at ICRA 2014 [32] . In this approach, each vehicle estimates and memorizes on-line the path of its predecessor as a set of points. After choosing a suitable position to aim for, the follower estimates on-line the predecessor's path curvature around the selected target. Then, based on a heuristic search, it computes an angular velocity using the estimated curvature. The optimization criteria used in this work allows the robot to follow its predecessor's path without oscillation while reducing the lateral and angular errors.

In october, Jano Yazbeck defended her Phd Thesis [2] .

Map Matching

Participant : François Charpillet.

This work [8] has been realized during the Intrade Projet with Maan Badaoui from Lille University. It addresses an important issue for intelligent transportation system, namely the ability of vehicles to safely and reliably localize themselves within an a priori known road map network. For this purpose, we have proposed an approach based on hybrid dynamic bayesian networks enabling to implement in a unified framework two of the most successful families of probabilistic model commonly used for localization: linear Kalman filters and Hidden Markov Models. The combination of these two models enables to manage and manipulate multi-hypotheses and multi-modality of observations characterizing Map Matching problems and it improves integrity approach. Another contribution of the paper is a chained-form state space representation of vehicle evolution which permits to deal with non-linearity of the used odometry model. Experimental results, using data from encoders’ sensors, a DGPS receiver and an accurate digital roadmap, illustrate the performance of this approach, especially in ambiguous situations.

Multi-Camera Tracking in Partially Observable Environment

Participants : Arsène Fansi Tchango, Olivier Buffet, Vincent Thomas, Alain Dutech.

Fabien Flacher (Thales ThereSIS) is an external collaborator.

In collaboration with Thales ThereSIS - SE&SIM Team (Synthetic Environment & Simulation), we focus on the problem of following the trajectories of several persons with the help of several controllable cameras. This problem is difficult since the set of cameras cannot simultaneously cover the whole environment, since some persons can be hidden by obstacles or by other persons, and since the behavior of each person is governed by internal variables which can only be inferred (such as his motivation or his hunger).

The approach we are working on is based on (1) the HMM (Hidden Markov Models) formalism to represent the state of the system (the persons and their internal states), (2) a simulator provided and developed by Thales ThereSIS, and (3) particle filtering approaches based on this simulator. Since activity and location depend on each other, we adopt a Simultaneous Tracking and Activity Recognition approach (STAR) as presented in current state-of-the-art approaches.

A first novelty lies in the use of a complex behavioral simulator. In a single-target setting, we demonstrated that it allows inferring the behavior of a complex individual, even in case of long periods of occlusions (when cameras do not cover the trajectory of the target). This idea led to publications in AAMAS-14 [16] , STAIRS-14 [18] , and ECAI-14 [17] .

A remaining issue is to find tractable algorithms for efficiently tracking multiple targets simultaneously, which requires using a factored particle filter (with one distribution per target). To that end, we use a Joint Probabilistic Data Association Filter with two key ingredients. The first ingredient is a particular model of dynamics that largely decouples the evolution of several targets, and turns out to be very natural to apply (which has led already to a publication in Fusion-14 [19] ). Then, the factorization a priori implies, for a given target, simulating each of its particles with each particle of each other target (which leads to a huge number of simulations). The second proposed ingredient is to simulate each particle of a given target only with a small number of “representatives” of each other target (and then, because more particles are produced than needed, a selection/resampling step is required).

Emergence et Developmental Learning

Participants : Alain Dutech, Matthieu Zimmer.

Yann Boniface (CORTEX, Loria) is an external collaborator

Following our ongoing work on using reinforcement learning for the control of redundant continous robotic systems, we explore how learning such complex tasks can benefit from a developmental approach, following some line of work already tested in robotics [50] .

“Emergence”, on of the key concepts grounding this work, has been presented – from an artificial intelligence perspective – and discussed with researchers from other fields. This lead to fruitful exchanges and a chapter in a bookdedicated to the dual aspects of (human gestures) : appearance and emergence [36] . “Developmental Learning” was also the main subject of a seminar in Lyon in which Alain Dutech has been invited [47] .

More concretely, we have developed several algorithms which mix artificial neural networks (like Dynamic Self-Organizing Maps or Reservoir Computing Network) with reinforcement learning mechanisms in order to build simple artificial systems that are autnomous and that learn without any exogeneous intervention from an external being. This work, initiated through two master thesis, is now the central topic ot the PhD of Matthieu Zimmer, started in october 2014.

Online Evolutionary Learning

Participants : Amine Boumaza, François Charpillet, Iñaki Fernandèz.

Evolutionary Robotics (ER) deals with the design of agent behaviors using artificial evolution. Within this framework, the problem of learning optimal decision functions (or controllers) is treated as a policy search problem in the parameterized space of candidate policies. In this work we are interested in learning optimal behaviors for swarm of mobile agents online (while solving the task). We adopt an online onboard distributed view  [56] , [48] and consider the learning process as executed at the agents' level in a decentralized way. This kind of algorithms raises several questions concerning the usefulness of selection pressure (partial views of population, noisy fitness values, etc.).

We studied the impact of task-driven selection pressures in on-line distributed ER for swarm behavior learning. We proposed a variant of the mEDEA  [45] algorithm in which we added a selection operator, in a task-driven scenario. We evaluated four selection methods that induce different intensity of selection pressure in a multi-robot navigation with obstacle avoidance task and a collective foraging task.

Experiments showed that a small intensity of selection pressure is sufficient to rapidly obtain good performances on the tasks at hand. We introduced different measures to compare the selection methods, and show that the higher the selection pressure, the better the performances obtained, especially for the more challenging food foraging task. This research was presented at the 13th International Conference on the Synthesis and Simulation of Living Systems [21] .

Frailty evaluation and Fall detection

Participants : Amandine Dubois, François Charpillet, Thomas Moinel, Maxime Rio.

This work is related to the IPL PAL and Satelor project and is related to Personal Assistant Living (PAL) for elderly people with loss of autonomy.

• Clinical evaluation of frailty in the elderly is the first step to decide the degree of assistance that elderly people require. No standard tests exist to detect the level of frailty, each clinician chooses his protocol among existing tools. There are clinical tests as Tinetti test, Timed Up and Go test for evaluating the degree of dependance and the frailty of elderly people. These tests consist in asking a person to realize exercises simulating movements of daily life. The physician evaluates the quality of gait and the balance of the patient. These tests are often used but, the disadvantage is that the final verdict relies primarily on a subjective opinion. The aim of our work is to provide new objective criteria to refine the elderly frailty quantification. We base ourselves on the frailty definition of Fried et al as being a clinical syndrome in which three or more of the following criteria are present: unintentional weight loss, self reported exhaustion, weakness (with regards to grip strength), slow walking speed and low physical activity. From this definition, we have defined two axis of development to evaluate the frailty of a person: Sensor based Activity recognition with the aim to follow and report daily life activities in order to detect evolution that coud reveal increased frailty [1] , gait analysis in order to assess gait pattern and their evolution over time [14] .

• An other PAL research domain, which is related to activity recognition, has attracted our attention: fall detection. Falls in the elderly is a major public health problem because of their frequency and their medical and social consequences. One of our objectives is to design an automatic system to detect fall at home, which in its final version will be made up of a network of RGB-D sensors, some of them being mobile embedded a wheel mobile robot.

The main contribution of this work has been to design a simple but robust method based on the identification and tracking of the center of mass of people evolving in an indoor environment through a RGB-D camera. Using a simple Hidden Markov Model whose observations are the position of the center of mass, its velocity and the general shape of the body, we have shown that we can surprisingly monitor the activity of a person with high accuracy, detect falls with very good accuracy without false positives and also measure some interesting parameter such as speed of gait, length of steps, etc. An experimental study, that is reported in [46] , has been driven in our smart apartment lab. 26 subjects were asked to perform a predefined scenario in which they realized a set of eight postures. 2 hours of video (216 000 frames) were recorded for the evaluation, half of it being used for the training of the model. The system detected the falls without false positives. This result encourages us to use this system in real situation for a better study of its efficiency. Therefore, we started this year an experimention in a room of a follow-up care and rehabilitation facility (OHS) in Nancy. "Office d’Hygiène Sociale" (OHS) is an association under the law of 1901. It supports nearly 800 people over 60 years and nearly 1,000 children and adults with disabilities. The association manages 26 facilities (40% health field, 40% medical-social field and 20% social field) and employs more than 1,500 professionals.

Posture recognition with a Depth camera

Participants : Abdallah Dib, François Charpillet, Xuan Nguyen, Alain Filbois [SED] .

In this research line, we focus our contribution on improving model-based approaches that use a population-based stochastic framework for full human body tracking using monocular depth camera. One of the major challenges in human tracking is the high-dimensional state spaces. To address this problem, we propose a tracking algorithm based on APF and CMA-ES. While APF has been widely applied for human tracking in RGB and depth images, the application of CMA-ES to human tracking is still limited. Yet, CMA-ES shares many similar ideas with APF and can be exploited to improve the performance of APF. Our key idea is to update the covariance matrix for sampling particles at each layer of APF, using a subset of best particles, an idea inspired from CMA-ES. The resulting algorithm is shown to greatly reduce the number of particles required for successful tracking. In the absence of image features such as texture or color, existing likelihood models for human tracking in depth images are often built by computing distances between data points and model points sampled on the surface of the human body model. When human body parts are close or when severe self-occlusions are present, these models fail to capture good pose hypotheses. As a result, existing approaches are unable to track a broad range of human motions. To deal with this issue, we propose a likelihood model which is based on comparing observed depth images and rendered depth images obtained by classic rendering techniques. Combining with our tracking algorithm, the proposed likelihood model has been shown to be effective when tracking under severe self-occlusions. To the best of our knowledge, our approach is the first model-based one that uses a population-based stochastic framework able to track full human body with non-frontal and unusual poses, using monocular depth camera.

Pressure sensing floor

Participants : Mihai Andries, François Charpillet, Olivier Simonin.

The use of floor-sensors in ambient intelligence contexts began in the late 1990's, with projects like ORL active floor, the Magic carpet by Paradiso et al., and the smart floor by Orr et al. These floors were, later on, integrated in smart environments, aimed at delivering assistance services like continuous diagnosis of users' health. According to the literature there are currently at least 6 main types of floor pressure sensing technologies: binary switches, piezoelectric, load cells, capacitive, polymer thick film (PTF), and photo interrupter sensors. Most of presented solutions extract a set of features for their tracking and identification task. Recently, sensing floors products like the SensFloor (a floor network of capacitive proximity sensors), Capfloor (a network of capacitive sensors), Elsi® smart floor (http://www.elsitechnologies.com ) and FloorInMotion (Tarkett France) started being commercialized by companies, mainly for the senior care industry.

We have ourselves developed a sensing floor. This load-sensing floor is composed of square tiles, each equipped with two ARM processors (Cortex m3 and a8), 4 load cells, and a wired connection to the four neighboring cells. Each tile has 16 light-emitting diodes which provide visual feedback. The processing units were manufactured by Hikob (http://www.hikob.com/ ). This prototype was originally designed as a medium of interaction for robots with distributed control, in an ant-like fashion. The computing unit available on each tile can register a virtual pheromone trace, that can then be transmitted to other robots, using either wired or wireless communication. In a different perspective, the sensing-floor acts merely as a sensor for an ambient intelligence. Using the magnetometer embedded on the processing unit of the tile, each tile can detect disturbances in its surrounding magnetic field, that can be caused by the presence of robots. Each tile also has an embedded accelerometer, that allows it to detect shocks that can be caused by objects or humans falling on the ground.

Several functionalities have been implemented this year on this prototype floor, including weight measurement, fall detection, footstep tracking and activity recognition. We also implemented heuristic real-time multi-user localisation (without user identification) in an indoor setting using this prototype floor.

Living assistant Robot

Participants : François Charpillet, Nicolas Beaufort, Abdallah Dib.

With LAR (living AssistanT Robot), a PIA projet which started in March, Abdallah Dib joined our team for a PhD. His work is about the development of a low cost navigation system for a robot evolving in an indoor environment. The main issue of his work is to design a Simultaneous Localisation and Mapping algorithm working in a dynamic environment in which people are moving. This is very challenging if we restrict the sensing capabilities of the robot with low cost sensors such as RGB-D camera. An important service we expect the robot to achieve, is realizing similar services as the one we described below: fall detection, activity recognition. This year first result have been published [11] . A feature based visual SLAM method that uses chamfer distance to estimate the camera motion from RGB-D images has been presented. The method does not require any matching which is an expensive operation and always generates false matching that affects the estimated camera motion. Our approach registers the input image iteratively by minimizing the distance between the feature points and the occupancy grid using a distance map. We demonstrate with real experiments the capability of the method to build accurate 3D map of the environment with a hand-held camera. While the system was mainly developed to work with RGB-D camera, occupancy grid representation gives the method the ability to work with various types of sensors, we show the capacity of the system to construct accurate 2D maps using telemeter data. We also discuss the similarities between the proposed approach and the traditional ICP algorithm.

Exploring an unknown environment with a team of mobile robots

Participants : François Charpillet, Olivier Simonin, Nassim Kaldé.

This work is the continuation of the work realized during the ANR Cart-O-matic (2010 to 2013). We address, here, the problem of efficient allocation of the navigational goals in the multi-robot exploration of unknown environment. Goal candidate locations are repeatedly determined during the exploration. Then, the assignment of the candidates to the robots is solved as the task-allocation problem. A more frequent decision-making may improve performance of the exploration, but in a practical deployment of the exploration strategies, the frequency depends on the computational complexity of the task-allocation algorithm and available computational resources. Therefore, this year, we have proposed an evaluation framework to study exploration strategies independently on the available computational resources. A comparison of the selected task-allocation algorithms deployed in multi-robot exploration has been done and published with Jan Faigl from Czech Technical University in Prague in the framework of the PHC project MACOREX.

An other point that is addressed by Nassim Kaldé is to consider the same problem but with dynamical environment in particular populated with human beings. First results of Nassim Kalde have been published in JFSMA'14 [33] . He published too the work done during his Master thesis [23] .