4 research outputs found
M.Thiriet, Editor TUMOR DETECTION IN MR LIVER IMAGES BY INTEGRATING EDGE AND REGION INFORMATION ∗
No full textAbstract. This paper describes a segmentation technique for 2D interventional MR images of liver tumours. Two features of MR data were likely to challenge existing segmentation methods. The first one is the inhomogeneous intratumoral texture, while the second one is the ”blurred ” appearance and the non-uniform sharpness of the tumour boundary. In order to detect the region of interest, we create the tumour contour map using a multithresholding technique and a measure of similarity between successive contours. Tumours presenting boundaries with non-uniform sharpness are segmented with an algorithm based on pixel aggregation and local textural information. Résumé. Cet article décrit une méthode précise et fiable de segmentation pour des images RM des tumeurs de foie. Les approches traditionnelles de segmentation d’images ne donnent pas de bons résultats dans ce cas-là, à cause de deux caractéristiques particulières des données d’entrée: l’inhomogénéité de la texture intratumorale et l’apparence floue de la paroi des tumeurs. Pour détecter la région d’intérêt, nous créons une carte des contours de la tumeur en utilisant une technique de multi-seuillage et une mesure de similarité entre les contours successifs. Les tumeurs qui présentent des contours flous de contraste variable sont détectées avec un algorithme basé sur l’agrégation des pixels et sur l’information locale de texture. 1
Computer Vision-Based Interface for the Control of Meta-Instruments
No full textAbstract. This paper describes a “virtual keyboard ” for the control of metainstruments. The proposed approach uses video input data and computer vision algorithms for tracking feet motion and their interaction with a planar keyboard with no force feedback. The design of the “virtual keyboard ” is directly inspired from the traditional, organ-style bank of foot pedals. The proposed approach accurately detects in real-time the hit of a keyboard with either one or both feet, as well as the location(s) of hit(s) (i.e. what keys have been “pressed”)
Monnet: Monitoring pedestrians with a network of loosely-coupled cameras
No full textMONNET is a visual surveillance system for tracking pedestrians over extended premises. The MONNET system is composed of intelligent nodes, which exchange information on the individually tracked pedestrians in an asynchronous manner. Each node in MONNET builds an appearance model for every observed pedestrian and compares it with models received from other nodes. The compact appearance models based on colour cues and face biometrics are stored locally on each node. The system is dynamically reconfigurable since its design allows for adding/removing nodes in a simple manner, comparable to the ‘plug and play ’ technology. MONNET also contains an optional ‘observer ’ node for interactive data visualization. This node displays a user interface which allows a human operator to observe and to interact in real-time with the distributed tracking process. MONNET was extensively tested with and without user input, and it is able to function correctly in both modes
