Christian Barat

Classification of indoor environment using only one ultrasonic sensor

We attempt to show that one ultrasonic sensor and one measure of a target is sufficient to classify indoor environment in four classes. We think the information in the echo even with electronic saturation can be extracted to perform the classification using statistical methods. We compare two statistical approaches (linear discriminant analysis and quadratic discriminant analysis). The feature extraction is also studied and the reduction of parameters is obtained with a sequential backward selection approach. Finally, we show qualitative performances on a real indoor environment.

Matching segments in stereoscopic vision

We have shown that its possible to realize a stereoscopic sensor with poor cameras. We developed image processing that is robust and allows us to quickly obtain results for the matching algorithm. We computed an important number of features on each segment, and with these features, we built 16-component vector used in the classification step. After an exhaustive study, we decided to combine two methods, Bayesian and neural, to construct an efficient classifier. The tests for indoor images had better than 90% good matching. With segment couples, it is possible to compute the 3D coordinates of the objects. Therefore, the mobile robot is able to localize and move about in the environment.

Ultrasonic multi-transducer processing for pattern recognition

This paper discusses the development of a new binaural ultrasonic sensor for mobile robot localisation and differentiation of simple objects, without environment scanning. The design of a recognition system requires careful attention to the following issues: definition of pattern classes, pattern representation, environment sensing, feature extraction and selection, classifier design and teaming. The recognition of objects (plane, corner, edge and cylinder) is achieved by processing of sonar signal using statistical methods (K nearest neighbours, linear and quadratic discriminant analysis), the parzen window method and neural networks which first identify and then exploit echo features: the frequency, slope, surface, length, amplitude and time-of-flight (TOF) defined as characteristics of these objects. In our study, several methods are used to extract the most discriminant features set, like sequential methods (Backward and Forward), optimal method (branch and bound). In addition, we use the principal component analysis (PCA) method to provide the correlation between the discriminant parameters.

Off-line localisation of a mobile robot using ultrasonic measurements

Regarding assistance to disabled people for object manipulation and carrying, the paper focuses on the localisation for mobile robot autonomy. In order to respect strong low-cost constraints, the perception system of the mobile robot uses sensors of low metrological quality, ultrasonic ring and odometry. That poses new problems for localisation, in particular. Among different localisation techniques, we present only off-line localisation. With poor perception means, it is necessary to introduce a priori knowledge on sensors and environment models. To solve the localisation problem, the ultrasonic image is segmented applying the Hough transform, well-adapted to ultrasonic sensor characteristics. The segments are then matched with the room, modelled and assumed to be rectangular. Several positions are found. A first sort, based on a cost function, reduces the possibilities. The remaining ambiguities are removed by a neural network which plays the part of a classifier detecting the door in the environment. Improvements of the method are proposed to take into account obstacles and non-rectangular room. Experimental results show that the localisation operates even with one obstacle.

Distinction between objects with ultrasonic biaural system and only amplitude

This study investigates the processing of sonar signal using statistical methods (quadratic and linear) for differentiation of commonly encountered features in indoor robot environments. Statistical method can differentiate two targets (Plane and Corner) with higher accuracy, improving on previously reported methods. It achieves this by exploiting identifying echo features: the differential frequency, length, amplitude and time-of-flight (TOF) characteristics of these targets. The features extraction is also studied and the reduction of parameters is obtained with a SBS (Sequential Backward Selection) approach. The results indicate that the amplitude information is more crucial than other features. The study suggests a wider use of statistical methods and amplitude information in sonar-based mobile robotics.

Neural and statistical classifiers. Can such approaches be complementary

Neural networks are efficient in certain pattern recognition sub-problems, especially in feature extraction and classification. In many cases neural and statistical techniques are seen as alternatives. Our aim is to verify if these approaches can give complementary responses in order to consider the implementation of fusion methods. The comparison is applied to three examples belonging to mobile robot localization: (i) laser range finder modeling, (ii) feature extraction from ultrasonic range finder data and (iii) localization by a stereoscopic camera. In each case the solution of the problem is based partly on a classifier. The paper compares the performances of a multilayer perceptron (MLP) known as an efficient classifier and three statistical methods-quadratic discriminant analysis (QDA), linear discriminant analysis (LDA) and Bayesian. The performances of the classifier are estimated by classical criteria such as success and misclassification percentages and the study is completed by a sharp analysis where the method results are crossed two by two to evaluate the success percentage of a method applied to the misclassified set of another one. Experiments show the set of patterns misclassified by the different classifiers does not completely overlap.

Toward a versatile ultrasonic sensor

The use of mobile robots, still limited to specific applications, can be applied to service robotics such as assistance for the disabled, if efficient yet low-cost perception systems are available. A multiaural and multimodal ultrasonic device seems to be an appropriate means for achieving such a difficult compromise. In a multimodal perception system, each operating mode is suited to a specific task. The paper presents three different main working modes based on the needs of a robot for obstacle avoidance, localization and target tracking. In order to take into account either a sensor anomaly or specific constraints imposed by the robot, variations of the main modes are proposed under the name reduced service mode (RSM). Related to the main mode which is optimal a RSM establishes a new compromise between performance and response time. The paper focuses on pattern recognition -notably for target tracking- using classification by neural network but above all by statistical methods. Results show that the ultrasonic received signal allows the extraction of complex objects from the environment with a good recognition percentage. Ultrasonic technology can be used for functions different from the only obstacle avoidance.

Modelling of a camera-3D range finder system

A laser range finder mounted on a site and azimuth turret is used as a 3D range camera. It forms, associated with a video camera, an original stereovision system. The internal structure of both images are the same but the resolution of 3D image stays low. By ignoring the acquiring speed of measures, spatial resolution is limited by the accuracy of deviation device and the laser footprint. The fact that the impact of the beam is not a point introduces spatial integration.To correct the average at depth discontinuities due to the beam footprint, a neural-network-based solution is reported.The use of such a multisensor system requires its calibration. As camera calibration is a well-known problem, the paper focuses on models and calibration methods of the range finder. Experimental results illustrate the quality of the calibration step in terms of accuracy and stability.The footprint correction is evaluated for both 1D and 2D range finder scannings.

Geometrical and physical models of a 3D range finder

According to the problem to be dealt with, the use of data acquisition system requires the knowledge of several models. In the case of the laser range-finder mounted on a site and azimuth turret, the geometrical model defines the co-ordinate transformation between measurement and reference frames. The paper shows the interest of a global calibration in comparison with a more classical approach which divides the problem into an internal model and external one. During the acquisition process another model called physical model may be useful. It takes into account the fact the laser beam impact is not at a point and introduces a spatial integration. The so-called footprint effect limits the lateral resolution of the range finder at depth discontinuities. In order to correct that effect, an inverse physical model, based on neural networks, is proposed. The improvement of the footprint correction is then evaluated for a 2D and 3D scanning.