Bruno Steux

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.

Hardware-friendly pedestrian detection and impact prediction

We present a system for pedestrian detection and impact prediction, from a frontal camera situated on a moving vehicle. The system combines together the output of several algorithms to form a reliable detection and positioning of pedestrians. One of the important contributions of this paper is a highly-efficient algorithm for classification of pedestrian images using a learned set of features, each feature based on a 5/spl times/5 pixels shape. The learning of the features is done using AdaBoost and genetic-like algorithms. The described application was developed as a part of the CAMELLIA project, thus all the algorithms used in this application are designed to use a special set of low level image processing operations provided by the smart imaging core developed in the project. Fusion of the various algorithms results and tracking of pedestrians is done using particle filtering, providing a good tool to predict the future movement of pedestrians, in order to estimate impact probability.

tinySLAM: A SLAM algorithm in less than 200 lines C-language program

This paper presents a Laser-SLAM algorithm which can be programmed in less than 200 lines C-language program. The first idea aimed to develop and implement a simple SLAM algorithm providing good performances without exceeding 200 lines in a C-language program. We use a robotic platform called MinesRover, a six wheels robot with several sensors. We based our work and calculations on a laser sensor and the odometry of the robot. The article presents the different capabilities of the platform and the way we use them in order to improve our programs. We also illustrates the difficulties encountered during the programming and testing of the algorithm. This work shows the possibility to perform complex tasks using simple and easily programmable algorithms.

A mathematical explanation via “intelligent” PID controllers of the strange ubiquity of PIDs

The ubiquity of PID controllers in the industry has remained mysterious until now. We provide here a mathematical explanation of this strange phenomenon by comparing their sampling with the the one of “intelligent” PID controllers, which were recently introduced. Some computer simulations nevertheless confirm the superiority of the new intelligent feedback design.

Yet Even Faster (YEF) real-time object detection

We present an improvement to the object detection scheme of Viola and Jones (2004), using the example of face detection. We present a new kind of visual features that is based on sampling individual pixels within the examined sub-window, instead of comparing the sums of pixel values in rectangular regions. Our features are faster and do not demand the preparation of an integral image, nor variance-normalisation of each sub-window. The result is an improved system that detects faces in quarter-PAL images in only 9 milliseconds per image on a 2.4-GHz pentium computer.

PUVAME - New French Approach for Vulnerable Road Users Safety

In France, about 33% of roads victims are VRU. In its 3rd framework, the French PREDIT includes VRU Safety. The PUVAME project was created to generate solutions to avoid collisions between VRU and bus in urban traffic. An important part of these collisions take place at intersection or bus stop. In this paper, we detail the hardware and software architecture designed and developed in the project. This solution is based on offboard cameras observing particular places (intersection and bus stop in our case) to detect and track VRU present in the environment. The position of the bus is also computed and a risk of collision between each VRU and the bus is determined. In case of high risk of collision, the bus driver is warned. The HMI to warn the bus driver is also described. Finally, some experimental results are presented

Robust real-time on-board vehicle tracking system using particles filter

Abstract We describe a system for detection and tracking of vehicles from a single on-board frontal camera, developed as a part of the European CAMELLIA project. Five image processing algorithms are used to detect target vehicles, classify them and maintain their exact localization. The fusion of the result of the algorithms is done using particle filtering. We assert that the particles filter forms the optimal mechanism to exploit the perceived data since it maintains the full probability density function based on all available algorithm in a given illumination and weather conditions. The algorithms are designed to exploit a set of low-level image processing operations, provided by a smart imaging core developed in the project. The result is that the system runs on 20 images/sec even on a regular pentium PC, and is design to run on real time using an ARM and the hardware core. The system was tested on many sequences and performs well even in hard conditions like rain and night.

Boosted algorithms for visual object detection on graphics processing units

Nowadays, the use of machine learning methods for visual object detection has become widespread. Those methods are robust. They require an important processing power and a high memory bandwidth which becomes a handicap for real-time applications. The recent evolution of commodity PC computer graphics boards (GPU) has the potential to accelerate those algorithms. In this paper, we present a novel use of graphics hardware for object detection in advanced computer vision applications. We implement a system for object-detection based on AdaBoost [1]. This system can be tuned to run partially or totally on the GPU. This system is evaluated with two face-detection applications. Those applications are based on the boosted cascade of classifiers: Multiple Layers Face Detection (MLFD), and Single Layer Face Detection (SLFD). We show that the SLFD implementation on GPU performs up to nine times faster than its CPU counterpart. The MLFD, in the other hand, can be accelerated using the (GPU) and performs up to three times faster than the CPU. To the best of our knowledge, this is the first attempt to implement a sliding window technique for visual object-detection on GPU, with promessing performance.

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.

A fast scan matching for grid-based laser SLAM using streaming SIMD extensions

The scan matching is one of the basic elements of several SLAM (Simultaneous Localization and Mapping) algorithms. There are a lot of researches that are interested in scan matching, but the majority deals with the algorithmic side only, without worrying about implementation tricks that can be very useful. This paper presents a simple and effective method to accelerate the scan matching step. Our method uses the computing power offered by the SSE instructions. These instructions, developed by Intel, allows the processing of several data simultaneously. This method does not require the use of a specific hardware (GPU for example), beceause SSE instructions exist in the majority of the commonly used processors (Intel, AMD). The examples and programs presented in the article are part of our SLAM algorithm, developed in the Robotics Center of Mines ParisTech.