Fabien Moutarde

Person re-identification in multi-camera system by signature based on interest point descriptors collected on short video sequences

We present and evaluate a person re-identification scheme for multi-camera surveillance system. Our approach uses matching of signatures based on interest-points descriptors collected on short video sequences. One of the originalities of our method is to accumulate interest points on several sufficiently time-spaced images during person tracking within each camera, in order to capture appearance variability. A first experimental evaluation conducted on a publicly available set of low-resolution videos in a commercial mall shows very promising inter-camera person re-identification performances (a precision of 82% for a recall of 78%). It should also be noted that our matching method is very fast: ~ 1/8s for re-identification of one target person among 10 previously seen persons, and a logarithmic dependence with the number of stored person models, making re- identification among hundreds of persons computationally feasible in less than ~ 1/5 second.

Precollapse scale invariance in gravitational instability

We numerically and analytically investigate the transition to the very nonlinear regime during the gravitational collapse of collisionless matter in an Ω=1 expanding universe. It is found numerically that the formation of halos from isolated and smooth initial overdensities leads to the progressive establishment, complete at the collapse time, of a power-law density profile. The evolutions from various types of initial perturbations are shown to all produce such power laws. The slopes are different, but the scale invariance itself appears quite generic

Robust on-vehicle real-time visual detection of American and European speed limit signs, with a modular Traffic Signs Recognition system

In this paper, we present robust visual speed limit signs detection and recognition systems for American and European signs. Both are variants of the same modular traffic signs recognition architecture, with a sign detection step based only on shape-detection (rectangles or circles), which makes our systems insensitive to color variability and quite robust to illumination variations. Instead of a global recognition, our system classifies (or rejects) the speed-limit sign candidates by segmenting potential digits inside them, and then applying a neural network digit recognition. This helps handling global sign variability, as long as digits are properly recognized. The global sign detection rate is around 90% for both (standard) U.S. and E.U. speed limit signs, with a misclassification rate below 1%, and not a single validated false alarm in >150 minutes of recorded videos. The system processes in real-time videos with images of 640times480 pixels, at ~20 frames/s on a standard 2.13 GHz dual-core laptop.

Priority-based coordination of autonomous and legacy vehicles at intersection

Recently, researchers have proposed various intersection management techniques that enable autonomous vehicles to cross the intersection without traffic lights or stop signs. In particular, a priority-based coordination system with provable collision-free and deadlock-free features has been presented. In this paper, we extend the priority-based approach to support legacy vehicles without compromising above-mentioned features. We make the hypothesis that legacy vehicles are able to keep a safe distance from their leading vehicles. Then we explore some special configurations of system that ensures the safe crossing of legacy vehicles. We implement the extended system in a realistic traffic simulator SUMO. Simulations are performed to demonstrate the safety of the system.

Improving pan-European speed-limit signs recognition with a new “global number segmentation” before digit recognition

In this paper, we present an improved European speed-limit sign recognition system based on an original ldquoglobal number segmentationrdquo (inside detected circles) before digit segmentation and recognition. The global speed-limit sign detection and correct recognition rate, currently evaluated on videos recorded on a mix of French and German roads, is around 94%, with a misclassification rate below 1%, and not a single validated false alarm in several hours of recorded videos. Our greyscale-based system is intrinsically insensitive to colour variability and quite robust to illumination variations, as shown by an on-road evaluation under bad weather conditions (cloudy and rainy) which yielded 84% good detection and recognition rate, and by a first night-time on-road evaluation with 75% correct detection rate. Due to recognition occurring at digit level, our system has the potential to be very easily extended to handle properly all variants of speed-limit signs from various European countries. Regarding computation load, videos with images of 640 times 480 pixels can be processed in real-time at ~20 frames/s on a standard 2.13 GHz dual-core laptop.

Spatial and temporal analysis of traffic states on large scale networks

We propose a set of methods aiming at extracting large scale features of road traffic, both spatial and temporal, based on local traffic indexes computed either from fixed sensors or floating car data. The approach relies on traditional data mining techniques like clustering or statistical analysis and is demonstrated on data artificially generated by the mesoscopic traffic simulator Metropolis. Results are compared to the output of another approach that we propose, based on the belief-propagation (BP) algorithm and an approximate Markov random field (MRF) encoding on the data. In particular, traffic patterns identified in the clustering analysis correspond in some sense to the fixed points obtained in the BP approach. The identification of latent macroscopic variables and their dynamical behavior is also obtained and the way to incorporate these in the MRF is discussed as well as the setting of a general approach for traffic reconstruction and prediction based on floating car data.

Analysis of Large-Scale Traffic Dynamics in an Urban Transportation Network Using Non-Negative Tensor Factorization

In this paper, we present our work on clustering and prediction of temporal evolution of global congestion configurations in a large-scale urban transportation network. Instead of looking into temporal variations of traffic flow states of individual links, we focus on temporal evolution of the complete spatial configuration of congestions over the network. In our work, we pursue to describe the typical temporal patterns of the global traffic states and achieve long-term prediction of the large-scale traffic evolution in a unified data-mining framework. To this end, we formulate this joint task using regularized Non-negative Tensor Factorization, which has been shown to be a useful analysis tool for spatio-temporal data sequences. Clustering and prediction are performed based on the compact tensor factorization results. The validity of the proposed spatio-temporal traffic data analysis method is shown on experiments using simulated realistic traffic data.

Decentralized model predictive control for smooth coordination of automated vehicles at intersection

We consider the problem of coordinating a set of automated vehicles at an intersection with no traffic light. The priority-based coordination framework is adopted to separate the problem into a priority assignment problem and a vehicle control problem under fixed priorities. This framework ensures good properties like safety (collision-free trajectories, brake-safe control) and liveness (no gridlock). We propose a decentralized Model Predictive Control (MPC) approach where vehicles solve local optimization problems in parallel, ensuring them to cross the intersection smoothly. The proposed decentralized MPC scheme considers the requirements of efficiency, comfort and fuel economy and ensures the smooth behaviors of vehicles. Moreover, it maintains the system-wide safety property of the priority-based framework. Simulations are performed to illustrate the benefits of our approach.

Analysis of network-level traffic states using locality preservative non-negative matrix factorization

In this paper, we propose to perform clustering and temporal prediction on network-level traffic states of large-scale traffic networks. Rather than analyzing dynamics of traffic states on individual links, we study overall spatial configurations of traffic states in the whole network and temporal dynamics of global traffic states. With our analysis, we can not only find out typical spatial patterns of global traffic states in daily traffic scenes, but also acquire long-term general predictions of the spatial patterns, which could be used as prior knowledge for modeling temporal behaviors of traffic flows. For this purpose, we use a locality preservation constraints based non-negative matrix factorization (LPNMF) to obtain a low-dimensional representation of network-level traffic states. Clustering and temporal prediction are then performed on the proposed compact representation. Experiments on realistic simulated traffic data are provided to check and illustrate the validity of our proposed approach.

COMBINING ADABOOST WITH A HILL-CLIMBING EVOLUTIONARY FEATURE SEARCH FOR EFFICIENT TRAINING OF PERFORMANT VISUAL OBJECT DETECTORS

This paper presents an efficient method for automatic training of performant visual object detectors, and its successful application to training of a back-view car detec- tor. Our method for training detectors is adaBoost applied to a very general family of visual features (called “control-point” features), with a specific feature-selection weak-learner: evo-HC, which is a hybrid of Hill-Climbing and evolutionary-search. Very good results are obtained for the car-detection application: 95% positive car detection rate with less than one false positive per image frame, computed on an independant validation video. It is also shown that our original hybrid evo-HC weak-learner allows to obtain detection performances that are unreachable in rea- sonable training time with a crude random search. Finally our method seems to be potentially efficient for training detectors of very different kinds of objects, as it was already previously shown to provide state-of-art performance for pedestrian-detection tasks.